TL;DR

• Best closed model: GPT-5.5. It scored 70%, ending its sibling’s six-board reign with a 4–2 head-to-head win over Sol and significant margins across the bottom half of the field.

• Cheapest closed model: Grok 4.5. For once, the cheap model is also on the podium: $0.18 per correct answer from the #2 seat, its first podium on this board.

• Best open-weights model: DeepSeek V4 Flash. On debut, it scored 55%, five points clear of the rest of the open field, at $0.013 per correct answer. Those are the cheapest results we have ever ranked this high.

• Real churn. Just 27 of 36 order calls between returning models repeated today (τ = 0.50). The crown changed hands, Grok climbed three seats, and Nemotron fell five.

Subscribe now

What to do with today’s board

Two debuts, one dethroning, and a board mostly defined by money.

DeepSeek V4 Flash just did what Luna did on Friday, but five seats higher: 55%, above Gemini and Opus 5, at a realized cost 17X cheaper than Sol’s. It is not fast. With a typical answer taking ~7 minutes. It’s a better fit for batch work than interactive loops. Within that lane, however, it is the best price per result option on the board.

Grok’s first podium comes with the same warning it always carries: the cheapest results in the top half, but one in eight attempts died to the clock.

The GPTs remain the safe pick, and today the order is GPT-5.5 first.

And if you were waiting for Qwen 3.8 Max to justify its price increase over the 3.7 model it replaces, today’s debut did not do it. The results were still mid table. The answers were faster. The bill was higher.

That is our read. Argue with it. Tomorrow’s board is the rebuttal.

Three stories lead today’s board. The rest are quick hits.

Subscribe now

1. The siblings trade the crown again

Sol’s six-board reign ended the way these things tend to end here: quietly, on the strength of a 4–2 head-to-head.

It is the first time in a week that either sibling has beaten the other by more than one problem.

GPT-5.5 took #1 at 70%. Sol’s fall to third says less about Sol, whose misses were again concentrated on problems nobody solved, than it does about today’s slate.

Our comparability controls read today’s problems as more difficult than Friday’s. Most of the field gave back points. The two models that did not, GPT-5.5 and Grok, moved past them.

The GPT pair has now occupied the top two seats on eight of the last nine boards and traded the crown twice in a week.

Every other vendor is competing for third.

2. Grok’s first podium, earned uphill

Grok 4.5’s ended today at #2, its first podium.

Grok finished level with Sol at 65% and moved ahead on the tiebreak. More importantly, it made the climb on a tougher slate, which makes the improvement harder to dismiss as noise.

It also remained the cheapest model in the top half of the table at $0.18 per correct answer.

The asterisk has not moved. One in eight attempts ran out the clock, and a typical answer takes nearly five minutes. Its wins arrive more slowly and less reliably than those of the models around it.

But a month ago, Grok was a budget row with reliability problems.

Today, it is a podium model with reliability problems, at roughly one-fifth the price of the models it just outscored.

3. DeepSeek does Luna’s trick, five seats higher

Friday’s story was a subpenny model solving 40% of the board.

Today, DeepSeek V4 Flash debuted at 55%, good for #5, above Gemini and Opus 5, at $0.013 per correct answer.

That is roughly double Luna’s solve rate and still just 1/17th of Sol’s cost, from a model that outscored half the board on its first morning.

The debut profile was clean:

• A full slate

• One clock loss

• A miss mix weighted toward problems the entire field found difficult

Its head to head results look like those of a real midtable contender. DeepSeek beat Nemotron 6–2, edged Gemini, split with the budget field, and lost clearly to the GPTs.

The catch is the stopwatch.

A typical answer takes about seven minutes, making it the second-slowest model on the board. For interactive work, that is likely disqualifying.

For anything that can run overnight, the cost comparison is not close.

The open weights crown changes hands on day one.

Quick hits

• Fable 5 held #4 at 60%, its seventh consecutive top-four finish. It still produces the board’s most expensive results at $1.04 per correct answer.

• Qwen 3.8 Max debuted at 48%, roughly where the 3.7 model it replaces tended to land. It was much faster, with a typical answer taking 331 seconds instead of its predecessor’s routine 700-plus, but one-sixth of its attempts still died to the clock. Better manners, same grades, bigger bill.

• Nemotron 3 Ultra fell five seats to last at 46%, surrendering its open-weights lead in one bad morning. The slate was clean, so the misses were real.

• Kimi K3 rebounded four seats to 50%, recovering most of yesterday’s collapse. Its four-board line, 52, 45, 36, 50, is the most volatile on the board.

• GLM-5.2 matched its best score in weeks at 50%, although another handful of provider errors cut its slate short.

• Gemini 3.6 Flash cooled to #6 at 52%. It remains the fastest model on the board, with a typical answer taking 82 seconds, but it is now flanked below by a debutant costing roughly one-fortieth as much and above by models it has not recently been able to catch.

Why the board moves at all

Just 27 of 36 order calls between returning models repeated today, producing a Kendall rank correlation of τ = 0.50.

Unusually, the churn reached the top of the table:

• A crown change

• A first ever podium

All three happened on the same morning.

Our comparability controls read today’s slate as more difficult than Friday’s, and most of the field’s scores fell with it. That makes the two models that improved anyway, GPT-5.5 and Grok, the day’s clearest signal.

When the same measurement produces a frozen board one week and a reshuffled board the next, both readings are doing their job. The quiet period showed that the ruler was steady. Today, that ruler caught three genuine moves at once.

Why not just trust the benchmarks you have already seen?

Public benchmarks eventually leak into training data. Scores inflate, and the benchmark begins measuring memorization instead of ability.

They are also snapshots. They may be accurate on the day they are published, but they become less representative as models, providers, and serving systems change.

We generate fresh, novel problems every day.

No model can have trained on a problem that did not exist until this morning. That is how a model that appeared four days ago can debut against the frontier on equal terms and take the open weights lead before its launch coverage has even cooled.

Today’s board used roughly $113 in compute over 14M tokens.

Subscribe, and each morning you will get the day’s best model, plus the fastest and cheapest way to get that performance, in your inbox before you start work.

Subscribe now

TL;DR

• Best closed model: GPT 5.6 Sol, on a tiebreak. Sol and GPT 5.5 finished dead level at 72% and split their head to head 1–1. Sol’s sixth straight crown rests on the strength of schedule math.

• Cheapest closed model: GPT 5.6 Luna, if 40% clears your bar. It delivered a correct answer for $0.007, the first sub penny figure we have ever printed. If 40% does not clear your bar, Terra solved half the board for a nickel per correct answer.

• Best open weights model: Nemotron 3 Ultra. It scored 56%, holding the open lead for a third straight day, this time on a nearly clean slate.

• Steady board. 41 of 45 order calls between returning models repeated, for a Kendall τ of 0.82. The entire returning top seven held its seat.

• Today’s tidbit: the two debutants are the two fastest models we have ever measured. Terra answered in a typical 54 seconds and Luna in 48, both beating the stopwatch record Gemini had owned since its debut.

  Subscribe now

What to do with today’s board

The top of the board says what it has said for two weeks: the GPT flagships first, Fable at a premium, and Gemini for speed.

What changed today is the bottom of your bill.

Terra solved half the board at $0.054 per correct answer, one quarter of Sol’s rate, and answered faster than anything we had ever timed until its sibling posted 48 seconds. For a high volume workload dominated by easy to medium calls, the play this week is a router:

Terra or Luna first. Escalate to Sol on failure. Let the expensive model touch only what the cheap one misses.

Luna’s 40% will not carry hard work, and its losing records against the entire top half say so plainly. But it does not need to. At less than a penny per correct answer, it only needs to catch the easy volume, and it did.

Kimi owners should read the quick hits.

That is our read. Argue with it. Tomorrow’s board is the rebuttal.

Two stories lead today’s board. The rest is quick hits.

1. The GPT ladder is now a full staircase

OpenAI put four models on today’s board, and they finished #1, #2, #8, and #11.

That sounds like a mixed day until you look at the price column:

• Sol: 72% at $0.22 per correct answer

• Terra: 50% at $0.054

• Luna: 40% at $0.007

Luna’s entire slate cost roughly what a coffee stirrer costs.

The debutants’ head to head records against the leaders are honestly bad. Terra went 0–9 against Sol, and Luna went 1–14. Neither model threatens anything above the middle of the table on capability.

That is not what they are for.

What they establish is a complete price ladder inside one vendor: full accuracy at a rate we already called the board’s best value, then roughly 70% of that accuracy at one-quarter of the price, then roughly 55% of it at one-thirtieth.

Nobody else on the board offers the middle and bottom rungs at anything close to those prices. Both new rungs also answer faster than every incumbent, including the model that previously held the speed record.

2. A dead heat at the top, decided by the schedule

Sol’s sixth straight crown is its narrowest.

Sol and GPT 5.5 both finished at 72%, split their head to head 1–1, and forced the ranking to the strength of schedule tiebreak, which weighs each model’s full record against the field.

Sol’s record was the sturdier one. It finished unbeaten against all eleven competitors and posted five statistically significant margins. GPT 5.5 dropped a handful more of the matchups beneath it.

That is about as thin as a crown gets.

It is also the tenth crown Sol has collected in just over two weeks. Sol and GPT 5.5 have now locked out the top two seats for seven consecutive boards, and nothing else today finished within seven percentage points of either one.

The anatomy of a penny

Luna’s 40% deserves a closer look because its failure profile is the healthiest kind.

It finished every problem. It lost nothing to the clock and nothing to its provider. It also answered faster than any model we have ever timed.

Every miss was simply a wrong answer.

Just over half of those misses came on problems the leaders solved, roughly the same mix seen among models charging forty times more in the same region of the table.

Compare that with Inkling’s July debut. Inkling also failed only by being wrong, but it scored 5%. Luna fails the honest way at eight times the accuracy and roughly one-hundredth of the price.

There is no hidden reliability tax here.

Luna is exactly what it says it is: a small, blisteringly fast model that gets six problems in ten wrong, priced like it knows it.

Quick hits

• Fable 5 took #3 at 65% on the Elo tiebreak over Gemini. It has now reached the podium six days straight, while remaining the board’s most expensive result at $0.92 per correct answer.

• Gemini 3.6 Flash matched Fable at 65% and lost the stopwatch title it had held since debut, twice in one morning, to models costing a fraction as much. It remains the fastest model in the top half at 85 seconds.

• Grok 4.5 climbed to 60%, its strongest board in two weeks, although one-sixth of its attempts still ran out the clock.

• Opus 5 held #7 at 52%, consistent with its emerging mid-table pattern. One-tenth of its attempts timed out. Its seven board line has settled into the low fifties, wobble included.

• Qwen 3.7 Max remained the slowest model at 48%. One-third of its attempts timed out, its worst reliability result of the month.

• GLM-5.2 posted 44% on yet another provider clipped slate, with one-fifth of its attempts lost before judging.

• Kimi K3 collapsed to last among the incumbents at 36%, its worst board ever. The slate was clean and the clock caused no trouble. The misses were real, and a majority came on problems the leaders solved. Four boards ago, Kimi cracked the top five. The budget tier gives and takes quickly.

Why the board moves at all

41 of 45 order calls between returning models repeated today, producing a Kendall τ of 0.82.

Every returning model in the top seven held its exact position. The only flips came from Kimi’s fall and the shuffle it caused beneath it.

Our comparability controls judged today’s slate level with Friday’s, and no model tripped the drift check.

A board this stable, on the same day two new models arrive, gives us the cleanest kind of debut reading. Terra and Luna entered a settled field, so where they landed is likely where they belong, not an artifact of a chaotic morning.

Why not just trust the benchmarks you have already seen?

Public benchmarks leak into future training runs. Scores inflate and gradually start measuring memorization instead of ability.

They are also snapshots: true on the day they are taken, and less true every day afterward.

We generate fresh, novel problems every day. No model can have trained on a problem that did not exist until this morning.

That is how we can price a brand new model’s actual output, including a correct answer for $0.007, on the same slate and the same morning as the champion it undercuts.

Today’s board consumed roughly $104 in compute over 13M tokens.

Subscribe, and each morning you’ll get the day’s best model, plus the fastest and cheapest way to reach that performance, in your inbox before you start work.

Subscribe now

TL;DR

• Best closed model: GPT-5.6 Sol. 72%, a fifth straight crown to close the month, with a winning record against all nine rivals.

• Cheapest closed model: also Sol. $0.21 per correct answer, the cheapest results on the board outright for the third day out of four. Grok’s discount sticker is within a cent, but its solve rate is 17 points lower.

• Best open weights model: Nemotron 3 Ultra. 53%, holding the open lead for a second day as Kimi continues to slide.

• Steady top, busy middle. 39 of 45 ordering calls between returning models repeated today (τ = 0.73). The top three seats have not moved in three days.

• Today’s tidbit: July’s crown ledger closed at nine apiece for Fable 5 and Sol, with four for GPT 5.5. Every one of Fable’s crowns came in the first half of July. Every one of Sol’s came after.

What to do with today’s board

Month end is a fair moment to zoom out.

The GPT pair has locked out the top two seats for six straight boards. Sol has been the best, and usually the cheapest, model on the table for two weeks. Nothing in today’s data argues against making it your default going into August.

The month’s other lesson is about launches. Two flagships debuted in July, and the board now has one of them oscillating around the middle of the table and the other off the roster entirely.

Opus 5 answered the question we left it yesterday, whether it could post a second strong board, with a six point retreat and a three seat slide. The second look it earned on Wednesday goes back on the shelf.

If you want Anthropic, the answer remains Fable, at a price.

That’s our read. Argue with it. Tomorrow’s board opens a new month to prove it wrong.

Two stories lead today’s board. The rest are quick hits.

1. The month the crown changed dynasties

Sol’s fifth straight crown closed July’s ledger at nine crowns for Fable, nine for Sol, and four for GPT 5.5.

The split could not be more lopsided in time. Fable’s reign belonged to the first half of the month, and it has not worn the crown since. Sol’s nine include the last five boards without interruption, four of them while also posting the cheapest results on the table.

Today was typical of the streak:

• 72% solve rate

• A winning record against every rival

• Significant margins over the budget half of the board

• 10–2 over Opus 5 (p = 0.04)

• A 5–4 squeaker over GPT-5.5, preserving the GPT one-two for a sixth straight day

Dynasties on this board have ended abruptly before. Fable’s ended the morning after it looked permanent.

But right now, the gap between Sol and everything not named GPT 5.5 is the most dependable fact we measure.

2. Opus 5 could not do it twice

Yesterday, we wrote that Opus 5’s first strong board needed a second one before it meant anything.

Today answered:

• 52%, down six points

• Three seats back to #7

• A statistically significant deficit to Sol again, 2–10 (p = 0.04)

The fair share of the drop is smaller than it looks, since our comparability controls read today’s slate as a few points stiffer for everyone. But even after that correction, Opus gave back real ground.

Its six-board line now reads:

57, 55, 45, 48, 59, 52

That looks less like a trajectory and more like a level: a model that belongs in the low fifties, wobbling on either side of it.

At $0.95 per correct answer, only its own stablemate is a worse deal on the board today.

Anthropic’s premium tier goes into August with Fable holding the podium and the newer flagship holding a question mark.

The anatomy of a 27%

GLM-5.2’s worst board of the month is a rare specimen: all three failure modes at once.

A fifth of its attempts ran out the clock. Its provider dropped or errored another fifth, shortening the slate it was even judged on. Of the misses it earned honestly, a majority were problems the leaders solved.

Any one of those is a bad day. All three stacked together is how a model that scored in the forties twice this week lands at 27%.

The prescription is different for each part, which is why we decompose the result instead of simply reading the number:

• The clock failures belong to the model.

• The capability misses belong to the model.

• The dropped slate belongs to the provider.

• None of the three belongs to the exam, because the same slate handed Sol its fifth straight crown.

Quick hits

• GPT-5.5 logged 70%, its sixth straight top-two finish. It has quietly finished within three points of Sol on four of the last five boards. The duopoly is two models, not one model and its shadow.

• Fable 5 held #3 at 62%, its fifth straight podium to close the month it began as champion. It remains the priciest model on the board, though its $0.97 per correct answer is its best rate in a week.

• Gemini 3.6 Flash reclaimed #4 at 57% and the stopwatch. Its typical latency was 73 seconds, fastest on the board for the sixth time in seven days.

• Grok 4.5 rebounded four seats to #5 at 55% and fixed its reliability overnight. Yesterday’s abandoned-attempt pile shrank to almost nothing. Its week reads 52, 48, 46, 55.

• Nemotron 3 Ultra held the open-weights lead at 53%, despite a provider hiccup trimming its slate around the edges.

• Qwen 3.7 Max slid to 48%, with a quarter of its attempts running out the clock. It remains the board’s slowest model at a typical 734 seconds.

• Kimi K3 extended its slide to 45%, three boards removed from its top-five breakthrough. Its week reads 52, 50, 49, 45: a slow leak, not a collapse.

Why the board moves at all

39 of 45 ordering calls between returning models repeated today (τ = 0.73).

For a third straight day, every flip happened below the podium:

• Grok rose four places.

• Gemini rose two.

• Opus 5 fell three.

Scores dipped a few points across most of the field, and our comparability controls attribute that decline to a slightly stiffer slate rather than deterioration in the models.

That is the mirror image of yesterday’s correction, applied in the opposite direction.

The symmetry is the point of measuring it at all. A friendlier morning should not manufacture a comeback story, and a stiffer one should not invent a collapse.

What survives both corrections this week is the same picture: two GPTs on top, a settled podium, and a second tier that reshuffles daily.

Why not just trust the benchmarks you’ve seen?

Public benchmarks leak into future training runs, so scores inflate and gradually start measuring memorization instead of ability.

They are also snapshots: accurate on the day they are taken, then less representative every day afterward.

We generate fresh, novel problems every day. No model can have trained on a problem that did not exist until this morning.

That is what makes a month-long ledger meaningful: a crown race that finished nine apiece at the top, measured across twenty-three different exams that no model saw twice.

Today’s board used approximately $107 in compute over 11.5M tokens.

Subscribe, and each morning you’ll get the day’s best model, plus the fastest and cheapest way to get that performance, in your inbox before you start work.

Subscribe now

TL;DR

• Best closed model: GPT-5.6 Sol. 74%, a fourth straight crown and an eighth overall, with a winning record against all nine rivals.

• Cheapest closed model: Grok 4.5, on the sticker. Its correct answers cost $0.18 each, but at a 46% solve rate, Sol remains the better buy. Its results cost three cents more and land 28 points higher.

• Best open-weights model: Nemotron 3 Ultra. 58%, a five-seat surge that takes the open lead back from Kimi after two days.

• Churny middle, steady ends. 34 of 45 order calls between returning models repeated today, with τ = 0.51. The top three stayed frozen while every position from 4th through 9th moved.

• Today’s tidbit: Sol did not miss a single problem that another model solved. Every problem it missed was failed by the entire board. Relative to the field, you cannot do better than that.

What to do with today’s board

The headline decision belongs to Anthropic.

Opus 5 posted its best board yet, finishing #4 at 59%, beating Gemini for the first time and escaping the significant deficits it carried throughout launch week. If you shelved it after its first few boards, today offers the first real argument for taking another look, with one important caveat discussed below.

Fable still beat it head-to-head, and both GPTs remain ahead on performance and price. Even on Opus 5’s best day, the pecking order holds: Sol first, GPT-5.5 second, then decide whether Anthropic loyalty is worth the premium.

Gemini remains the speed pick. Its fall to sixth still came with the fastest answers on the board, a typical 75 seconds.

Grok owns the cheapest sticker, but it also had its worst reliability day of the month. Treat that $0.18 as a sale price on a model that did not show up today.

That is our read. Argue with it. Tomorrow’s board is the rebuttal.

Two stories lead today’s board. The rest are quick hits.

1. Opus 5’s best board, with an asterisk it half-earns

Five boards: 57%, 55%, 45%, 48%, and now 59%.

The jump is real on the scoreboard. Opus climbed three seats to #4, finished above Gemini for the first time, and erased the statistically significant deficits to Sol and GPT-5.5 that defined its week. Today’s head-to-head records were 1-7 against Sol and 3-8 against GPT-5.5, neither statistically significant.

The asterisk is slate difficulty.

Our comparability controls indicate that today’s problems were slightly friendlier than yesterday’s, and most of the field rose with them. Opus 5’s improvement exceeded the field-wide lift, which is the part it earned.

What it has not earned yet is a flagship’s place in the hierarchy.

Fable still won their head-to-head 4-2. Both GPTs still sit above Opus on results while delivering correct answers at roughly half the realized price. One strong board after four soft ones is still only one data point.

The trendline argument we made two days ago cuts both ways. Now Opus needs a second board like this one.

2. Four crowns in a row, and a perfect miss profile

Sol’s 74% came with something we have not measured before: not one of its misses was a problem another model solved.

The entire board went zero-for on every problem Sol dropped.

Otherwise, its sweep of the field looked familiar. Sol posted winning records against all nine rivals, including significant margins over the budget half of the board:

• 13-2 over Grok, with p = 0.007

• 10-0 over Kimi

GPT-5.5 kept the race close again, losing the direct comparison 3-2, and quietly logged another strong day at 72%.

The pair have now locked out the top two seats for five straight days. The GPT duopoly has become the most stable fact on the board.

The anatomy of a recovery

Opus 5’s jump decomposes cleanly.

Its clock problem stayed fixed. For a second straight day, barely any attempts were lost to the deadline, following the reliability wobble earlier in the week.

Its wrong answers also improved. Yesterday, roughly half of its misses were on problems that the leaders solved. Today, that number fell to about a third. In other words, the problems Opus missed were more often the problems that stumped everyone. That is the signature of a stronger day, not merely a luckier one.

Its speed held too, at a typical 123 seconds, second-fastest among the leaders.

Grok’s collapse is the same decomposition in reverse. Nearly a fifth of its attempts died to the clock, and a majority of its misses came on problems the leaders solved.

One model closed both gaps at once. The other opened both.

Quick hits

• Fable 5 held #3 at 64%, earning its fourth straight podium. It remains the board’s most expensive model at $1.28 per correct answer, roughly six times Sol’s rate.

• Nemotron 3 Ultra surged five seats to #5 at 58%, its best board this month, and reclaimed the open-weights lead. Its recurring provider errors barely affected today’s run.

• Qwen 3.7 Max matched its best score of the month at 56%, while remaining the board’s slowest model at a typical 863 seconds. Nearly a quarter of its attempts ran out the clock.

• Kimi K3 gave back its rebound, falling three seats to #8 at 49%. Two strong boards followed by one soft one leave the open-weights race with Nemotron looking like a week-to-week coin flip.

• Grok 4.5 fell to #9 at 46%, with its worst reliability day of the month. The cheapest sticker on the board also bought the most abandoned attempts.

• GLM-5.2 brought up the rear at 42%, on another slate clipped by provider errors and carrying another asterisk.

• Roster note: Inkling sat out today’s board after three consecutive finishes in the basement.

Why the board moves at all

34 of 45 order calls between returning models repeated today, producing τ = 0.51.

The churn was concentrated entirely in the middle:

• Nemotron rose five seats.

• Opus 5 rose three.

• Kimi fell three.

• Grok fell three.

The top three positions did not move for a second straight day.

Scores increased across most of the field at once, and our comparability controls attribute most of that lift to the slate rather than the models. That correction is essential before interpreting Opus 5’s jump, and it is why we apply it.

A churny middle beneath a frozen top is the board sorting its second tier, not questioning its first.

Why not just trust the benchmarks you’ve seen?

Public benchmarks leak into future training runs. Scores inflate, and the benchmark gradually measures memorization instead of ability.

They are also snapshots: accurate on the day they are taken, and increasingly wrong every day afterward.

We generate fresh, novel problems every day. No model can have trained on a problem that did not exist until this morning. When a day’s slate is friendlier or more difficult than usual, we measure that too and correct for it before crediting a model’s jump.

Today’s board required $110.91 in compute over 12M tokens.

Subscribe, and each morning you’ll get the day’s best model, plus the fastest and cheapest way to reach that performance, in your inbox before you start work.

Subscribe now

TL;DR

• Best closed model: GPT-5.6 Sol. A 70% solve rate delivered its third consecutive crown and seventh overall. Sol posted a winning record against all ten rivals, with statistically significant margins over seven.

• Cheapest closed model: also Sol, again. At $0.21 per correct answer, Sol produced the board’s cheapest results for a second straight day. Grok finished two cents behind while solving far fewer problems.

• Best open-weights model: Kimi K3. A 50% solve rate secured the open lead for a second day and put Kimi in the top five for the first time.

• The calmest board we have measured. 51 of 55 order calls between returning models repeated, producing τ = 0.85. The top four seats did not move at all.

• Today’s tidbit: Inkling has doubled its debut score across three boards, from 5% to 10%. At this rate, it reaches the middle of the table sometime in September. It is still last and still lost to every model on the board.

What to do with today’s board

Nothing changed at the top, which is itself the finding.

After a week of dethronings, debuts, and collapses, the market finally gave you a stable answer:

• Sol is the pick on both results and price.

• GPT-5.5 is the cautious second choice.

• Fable 5 is the Anthropic pick, at a substantial premium.

• Gemini 3.6 Flash is the speed pick, one notch down on capability.

If you spent the last week waiting for the churn to settle before committing a workload, this is what settled looks like.

The quiet board also sharpens one decision. Opus 5 arrested its slide at 48%, but four boards in, it has yet to finish above its own house’s alternative. It also continues to lose to both GPTs by significant margins.

Waiting for board five is no longer caution. It is a choice.

That’s our read. Argue with it. Tomorrow’s board is the rebuttal.

Two stories lead today’s board. The rest are quick hits.

1. Seven crowns, and the misses that prove it

Sol’s third straight crown came at 70%, accompanied by the usual sweep: winning records against all ten rivals and statistically significant margins against seven.

That included:

• 10–1 over Opus 5 (p = 0.01)

• 24–0 over Inkling

The sharper evidence, however, is in what the leaders got wrong.

Sol and GPT-5.5 each missed only a handful of problems today, and those misses were almost entirely problems that nobody on the board solved. The two GPTs are not dropping problems the field can handle. They are being stopped only where everyone is stopped.

Contrast that with the bottom half of the table, where at least a third of each model’s misses came on problems the leaders successfully cleared.

2. Opus 5 stops falling, which is not the same as recovering

Board four: 48%. One seat gained. Slide arrested.

The good news for Anthropic’s newest flagship ends there.

Opus 5 still:

• Lost 1–10 to Sol, a significant margin

• Lost 1–9 to GPT-5.5, also significant

• Went 0–5 against Fable 5 (p = 0.06)

Its four-board line now reads:

57% → 55% → 45% → 48%

That looks less like a temporary stumble and more like a model that has found its level. That level is the middle of the table.

Opus remains quick, with a typical solve time of 129 seconds, and it reduced its clock losses from yesterday. But it is priced against the flagships while solving against the middle of the field.

One seat of recovery on the calmest day of the week is thin evidence of a turnaround.

What would change the story is a board above Fable. Opus 5 has not come close to one yet.

Quick hits

GPT-5.5 remains the closest challenger

GPT-5.5 held #2 at 68%, recording its fourth top-two finish in five boards. The sibling gap was 1–2 today, about as close to a tie as the math allows.

Fable holds its seat, but not its value

Fable 5 slipped to 60% but retained third place. Its results remain the most expensive on the board at $1.27 per correct answer, roughly six times Sol’s rate.

Gemini still owns the clock

Gemini 3.6 Flash held #4 at 57% and remained the board’s fastest model, with a typical solve time of 75 seconds.

Five boards in, Gemini’s latency floor looks like everyone else’s ceiling.

Kimi K3 confirms the rebound

Kimi K3 reached 50%, climbed one seat to fifth, and recorded its first top-five finish. It also retained the open-weights lead for a second consecutive day.

Qwen keeps the slowest-model title

Qwen 3.7 Max held at 45%, with roughly one-fifth of its attempts running out the clock.

GLM improves, with another provider asterisk

GLM-5.2 improved to 43% on another provider-shortened slate. It was the model’s second-highest score this month, once again accompanied by an asterisk it did not earn itself.

Inkling reaches double digits

Inkling climbed to 10% on its third board, its best result yet. It still posted a losing record against every model and remains farther from the leaders than anything else we track.

Why the board moves at all

Today, it mostly did not.

51 of 55 order calls between returning models repeated, producing τ = 0.85, the highest stability we have recorded. The only four flips were single-seat swaps in the middle of the table.

Scores drifted down by a few points across nearly the entire field. Our comparability controls attribute that dip to the models and their serving conditions rather than to a harder problem set. It is the same quiet variance that moves every model slightly on any given day.

A stable board is what makes the volatile ones meaningful.

When the same measurement produces a four-way tie one week and a frozen top four the next, you are watching the models change, not the ruler.

Why not just trust the benchmarks you’ve seen?

Public benchmarks leak into subsequent training runs. Scores inflate and gradually begin measuring memorization instead of ability.

They are also snapshots: true on the day they are taken, and less representative every day afterward.

We generate fresh, novel problems every day. No model can have trained on a problem that did not exist until this morning.

That is why a stable week means something here. The order held across five consecutive sets of problems that no model had ever seen, not five reruns of the same exam.

Today’s board consumed $114.77 in compute.

Subscribe, and each morning you’ll get the day’s best model, plus the fastest and cheapest way to achieve that performance, in your inbox before you start work.

Subscribe now

TL;DR

• Best closed model: GPT-5.6 Sol. 75% for a second straight day and a sixth crown overall, with a winning record against every rival and statistical separation from more than half the board.

• Cheapest closed model: also Sol. $0.21 per correct answer, the cheapest result on the entire board. The best model is currently also the best deal, which does not happen often.

• Best open-weights model: Kimi K3. 52%, a five seat rebound from its worst board, taking the open lead from Nemotron.

• Steady up top, churn below. 48 of 55 order calls between returning models repeated, for a Kendall τ of 0.75, with most of the movement concentrated in the middle.

• Today’s tidbit: Opus 5’s slide has a trendline now. Three boards: 57%, 55%, 45%. Today, its deficits to both GPTs and its own stablemate Fable became statistically significant, including a 0–9 blanking by Fable.

What to do with today’s board

Sol needs no caveats today: best results, cheapest results, near fastest, and six crowns in nine boards.

If you are running anything else for this class of work, the burden of proof has flipped to the anything else.

The Anthropic story has inverted. Fable 5 is the house’s standard-bearer at #3 with a 68% solve rate, and it beat the newer, cheaper Opus 5 nine problems to zero. If you migrated from Opus 4.8 to Opus 5 during launch week, these three boards say to migrate back or sideways to Fable.

Kimi K3 is the quiet value note: back to 52% at $0.24 per correct answer, making it the best open-weights buy today.

And Inkling’s second board, at 8%, confirms its first.

Two stories lead today’s board. The rest are quick hits.

1. Six crowns, and the cheapest seats in the house

Sol repeated at 75% and separated further from the field: a winning record against all ten rivals and statistically significant margins over seven of them.

Those margins ranged from 13–1 against Opus 5, with p = 0.002, down through the budget half of the board.

Against its closest peers, the gaps stayed friendly but consistent:

• 5–2 over Fable 5

• 4–2 over GPT-5.5

• 7–2 over Gemini 3.6 Flash

What makes this crown different from the five before it is the price column.

At $0.21 per correct answer, Sol delivered the cheapest results on the board, beating models that solved barely half as much.

When the accuracy leader is also the value leader, the usual capability-versus-cost tradeoff disappears. The Pareto chart below is one dot followed by excuses.

2. Opus 5’s trendline points the wrong way

Board one: 57%, soft debut.
Board two: 55%, benefit of the doubt.
Board three: 45%, and the doubt is gone.

The deficits that were merely suggestive are now statistically significant:

• 1–13 against Sol, p = 0.002

• 2–12 against GPT-5.5, p = 0.01

• 0–9 against Fable 5, p = 0.004

That Fable result marks the first time this month one flagship has blanked another.

Our comparability controls read today’s slate as level with yesterday’s, and no model tripped the drift check. This is therefore a story about the model and its serving, not the exam.

With Opus 4.8 off today’s roster, the awkward direct comparison is gone, but the conclusion is not: Anthropic’s newest model is currently its weakest ranked offering, at a price that only undercuts its own stablemates.

Three boards is a pattern. Tomorrow tells us whether it holds.

The anatomy of a 45%

Opus 5’s slide decomposes into two failure modes, and neither flatters it.

First, straightforward wrong answers. Half of its misses came on problems that a majority of the leaders solved, up from its debut. The gap is not the slate getting harder.

Second, a new clock problem. A tenth of its attempts ran out the time budget today, its worst showing there yet, even though its completed answers stayed fast at a typical 105 seconds.

Fast when it finishes, wrong too often when it does, and now not always finishing. That is a capability problem with a reliability problem stacked on top.

Contrast that with Kimi K3, whose rebound came the honest way: one lost attempt all day, quick answers, and misses concentrated on the problems everyone found difficult.

Quick hits

• GPT-5.5 climbed back to #2 at 70%, restoring the GPT one-two. The sibling gap remained a coin flip at 2–4, but the pair has now owned the top of the board for two consecutive days.

• Fable 5 added ten points across two boards to reach #3 at 68%, its best finish in a week, though it still produced the priciest results at $0.98 per correct answer.

• Gemini 3.6 Flash cooled to #4 at 62%, while remaining the fastest model on the board at a typical 79 seconds. This looks more like regression than a slide; its four-board average still beats everything below the GPTs.

• Grok 4.5 reached #5 at 55%, but one in eight attempts ran out the clock, its worst reliability day this month.

• Qwen 3.7 Max matched Kimi at 52%, yet remained the board’s slowest model at a typical 899 seconds, with nearly a quarter of its attempts dying to the clock.

• GLM-5.2 fell to 32%, with a third of its slate lost to the clock rather than wrong answers. Different failure, same unusable result.

• Inkling improved to 8% on board two. It also picked up a new habit: one-sixth of its attempts now run out the clock, which did not happen during its debut. Two boards, no case for a workload yet.

Why the board moves at all

48 of 55 order calls between returning models repeated today, producing a Kendall τ of 0.75.

The seven flipped comparisons live almost entirely in the middle of the table:

• Kimi rebounded five seats.

• GPT-5.5 and Fable each climbed two.

• Gemini and Opus 5 gave seats back.

The top seat did not move. Neither did the bottom one.

Our comparability controls read the slate as level with yesterday’s, and no model tripped the drift check. The mid-table churn therefore looks like serving variation and ordinary day-to-day variance doing what they do.

Meanwhile, the two ends of the board, a champion pulling away and a debutant confirming its floor, are about as stable as anything we measure.

That mix, stable extremes and a restless middle, is what most mornings actually look like.

Why not just trust the benchmarks you’ve seen?

Public benchmarks leak into the next training run, causing scores to inflate until they begin measuring memorization instead of ability.

They are also snapshots: true on the day they are taken, and less true every day afterward.

We generate fresh, novel problems every day.

No model can have trained on a problem that did not exist until this morning. That is how a three board trendline on a brand new flagship can exist at all: the same measurement process, fresh material, three mornings in a row.

Today’s board consumed $102.17 in compute.

Subscribe, and each morning you’ll get the day’s best model, plus the fastest and cheapest way to get that performance, in your inbox before you start work.

Subscribe now

TL;DR

• Best closed model: GPT-5.6 Sol. 75%, back on top after two days off the throne, and this time it beat every other model head-to-head, the first board in weeks with real separation at #1.

• Cheapest closed model: Grok 4.5. $0.19 per correct answer at 52%, edging Sol and Gemini (both $0.21). At a two-cent spread, the crown is still the better buy.

• Best open-weights model: Nemotron 3 Ultra. 42%, holding the open lead for a fourth board despite a rough day for the whole open field.

• Today’s tidbit: Thinking Machines’ Inkling debuted at 5%, a twentieth of the champion’s rate, against every rival’s double digits. It was fast and cheap about it, which does not help.

What to do with today’s board

Sol is the pick today without an asterisk: highest solve rate, a winning record against all twelve rivals, near-cheapest results, and quick. If you standardized on it during the four-crown streak, nothing that happened since should move you.

Gemini remains the value alternative one notch down, same $0.21 per correct answer, fastest of the leaders.

On Opus 5, the second board says what the first did, just softer: it climbed to #6 at 55%, still below Opus 4.8, still losing the direct matchup to Sol 2-10. Anthropic loyalists should stay on Opus 4.8 for another board at least.

And give Inkling a wide berth: at a 5% solve rate, the 93% discount buys you answers you cannot use.

Three stories lead today’s board; the rest is quick hits.

1. Sol’s crown, this time with a gap

Three days ago the top of the board was a four-way tie and GPT-5.5 held the crown by a single problem.

Today Sol took it back the emphatic way: 75%, ten points clear of second place, with a winning record against all twelve rivals.

The margins against its peers were not coin-flips this time: 7-1 over Fable (p=0.07), 6-1 over Opus 4.8, 7-2 over its sibling GPT-5.5, 10-2 over Opus 5 (p=0.04), and clean sweeps of the budget field topped by 28-0 over the debutant.

We have measured five Sol crowns now, and this is the first where the gap to the rest of the leaders looks like a gap rather than noise. Whether it holds is tomorrow’s question; today the board had one best model by a distance.

Share

2. Inkling debuts at 5%

Thinking Machines’ first board is the worst debut we have measured: 5%, with a losing record against every model on the table, including 2-11 against Muse Spark, the board’s perennial basement.

The strange part is how healthy the failure looks. It answered every problem, quickly, a typical 80 seconds, among the fastest, with substantial, confident output every time. The answers were simply wrong, almost every time.

And despite the bargain sticker, at $0.94 per correct answer its results cost more than four times Sol’s.

Nothing about the day offers it an excuse, either: it got a full slate, healthy serving, and plenty of clock, and on problems it had never seen it could not hang with a single incumbent.

You will read a very different account of this model elsewhere this week. We can only tell you what it solved.

The anatomy of a 5%

There are three ways to fail on this board:

1. Wrong answers

2. The clock

3. Your provider

Today produced a clean specimen of each, and they deserve different reactions.

Inkling: the model failed

Inkling’s failure is the first kind, undiluted.

It finished everything it started, quickly, and got nearly all of it wrong. Just under two thirds of its misses were problems a majority of the leaders solved.

That pattern points at the model, not the infrastructure.

Qwen 3.7 Max: the clock won

Qwen 3.7 Max is the second kind.

When it actually finished, it passed nearly two thirds of the time. But almost a third of its attempts ran out the clock, and those count against it.

Give it a workload without a deadline and it is a different model.

GLM-5.2: the provider failed

GLM-5.2 is the third kind.

Its 40% came on a slate its provider’s errors cut short, so the number describes the serving as much as the model.

One of these three earns an avoid, one earns a use-with-patience, and one earns a check-back-tomorrow.

The scoreboard alone will not tell you which is which; the failure mode does.

3. Opus 5, board two: better, still behind its predecessor

Opus 5 climbed two seats to #6 at 55%, up from its soft debut.

The direction is right and some of the detail is encouraging: it was quick again, a typical 107 seconds, and its results improved where the field’s mostly slipped.

But the two comparisons that define its debut have not flipped. It still trails Opus 4.8, 62% to 55%, and lost the direct matchup 5-8. And against the champion it went 2-10, one of only two significant separations Sol recorded among the leaders.

Two boards in, the pattern reads: a faster, somewhat cheaper Opus that solves less than the one it replaces. If that is the trade Anthropic intended, the price cut is doing the arguing.

Quick hits

• Gemini 3.6 Flash held #2 at 65%, fastest of the leaders again and tied with Sol on cost per correct answer. Four boards in: one wobble, three strong finishes. It has earned the default-alternative slot.

• GPT-5.5 gave back the crown with a ten-point slide to 62%, and lost the sibling matchup 2-7. Its day-one dominance is now a one-day story.

• Fable 5 parked at #5 (60%) again, mid-pack results at nearly the priciest rate on the board. The case for its premium shrinks by the day.

• Kimi K3 had the roughest day of the returning field, sliding to 35%, its worst board yet.

• GLM-5.2 scored 40% on the sharply shortened slate its provider’s errors left it; read that number with the asterisk it comes with.

• Qwen 3.7 Max returned to the board at 45% and immediately reclaimed its old title: slowest model measured, a typical answer at 653 seconds.

• Muse Spark 1.1 logged 30% on its seventh straight incomplete slate.

Why the board moves at all

37 of 45 order calls between returning models repeated today (τ = 0.64), and the flips sit almost entirely below the podium: Kimi and Nemotron sliding, Opus 5 climbing, GPT-5.5 dropping out of the tie at the top.

Our comparability controls read today’s slate as statistically level with the last board, and no model tripped the drift check, so the swings, Sol’s surge included, are the models and their serving rather than the problems.

Worth naming: a 75% day and a 5% day landed on the same slate, on the same morning, which is the widest spread between best and worst we have ever measured on identical fresh problems.

Why not just trust the benchmarks you’ve seen?

Public benchmarks leak into the next training run, so scores inflate and start measuring memorization instead of ability. They’re snapshots: true the day they’re taken, wronger every day after.

We generate fresh, novel problems every day. No model can have trained on a problem that didn’t exist until this morning, which is how a lab’s first model can be measured against the frontier the week it ships and land at 5% while the launch coverage says otherwise.

(Today’s board took about $137 in compute.)

Subscribe, and each morning you’ll get the day’s best model, plus the fastest and cheapest way to get that performance, in your inbox before you start work.

Subscribe now

Today’s board cost $147.78 in compute over 12.3M tokens. So listen up.

TL;DR

• Best closed model: GPT-5.5. 75%, a new #1 at $0.40 per correct answer, though the lead is a coin-flip: 4-3 over both Sol and Gemini, neither separation significant.

• Cheapest closed model: Gemini 3.6 Flash. $0.195 per correct answer at 72%, the cheapest results on the whole board from a model sitting second. It is the value pick.

• Best open-weights model: Nemotron 3 Ultra. 57%, holding the open lead for a third day.

• Real reshuffle. Only 27 of 36 order calls between returning models repeated (τ = 0.50): Sol’s four-day reign ended and Gemini climbed four seats in a day.

• Today’s tidbit: the new flagship landed soft. Opus 5 debuts at #8 (57%), well below Opus 4.8 at #4 (72%), and it lost to its own predecessor 3-8.

What to do with today’s board

The top of the board is a four way tie: GPT-5.5, Gemini, Sol, and Opus 4.8.

So pick on price and speed, not the ranking.

Gemini is the clear value call, cheapest results on the board and second-fastest, sitting one point off the top. GPT-5.5 is the nominal #1 if you want the highest raw number.

The one thing not to do today is rush to Opus 5. On its first board it is slower on results than the Opus it replaces, more expensive per correct answer than everything except the two Anthropic flagships, and it lost head-to-head to Opus 4.8.

Stay on Opus 4.8, or move to the GPTs or Gemini, until Opus 5 proves a second board.

And Fable’s premium is still the hardest to defend on the table: #5 at $1.24 per correct answer.

Three stories lead today’s board; the rest are quick hits.

Subscribe now

1. Opus 5 arrives, and Opus 4.8 beats it

The headline launch of the week landed at #8.

Opus 5 solved 57% on its first board, and the model directly above it in the lineage, Opus 4.8, solved 72% and beat it head-to-head 8-3.

Against the top of the table it was worse: 2-9 to Sol, GPT-5.5, and Gemini alike, each at p=0.065, which is as close to a real gap as one board can show.

The new model is not slow: a typical answer took 88 seconds, among the quickest on the board. But three of its attempts ran out the clock, and the answers it did return cost $0.80 per correct one, the third-priciest results on the table.

A debut this soft from a top line flagship is the kind of thing we exist to catch, because you would not learn it from the launch post.

It’s one board, and the model may climb. Today it sits in the bottom third, under the model it was built to replace.

2. GPT 5.5 is today’s best

Sol’s streak ended at four.

GPT-5.5 moved up to #1 at 75%, but calling it a lead overstates things. It went 4-3 over Sol, 4-3 over Gemini, and 5-3 over Opus 4.8, and not one of those is significant.

The four best models are inside three points of each other, tighter than yesterday’s top.

Where GPT-5.5 does pull clear is the field below: 8-1 over Grok (p=0.04), 11-0 over Kimi, 16-0 over Muse, all decisive.

So the useful way to read today’s #1 is not “GPT-5.5 is the best model” but “four models are effectively tied at the top and GPT-5.5 has the cleanest record against everything beneath them.”

Sol did nothing wrong; it solved 72% and lost the crown by a single problem to a sibling it splits evenly.

3. Gemini’s redemption arc

Two days ago Gemini debuted at #3 and then slid to #6, and the read here was to give it a third board before judging.

The third board is in: #2 at 72%, up four seats, tying both GPTs head-to-head and beating Fable 6-2.

It did it while staying the cheapest and second-fastest model on the table, $0.195 per correct answer at a 65-second typical answer.

That combination, top of table accuracy at the bottom of the price and latency charts, is the strongest value case on the board right now.

The debut wobble looks like noise; the model looks real.

Quick hits

• Fable 5 slid to #5 at 62%, down three seats, still the priciest results on the board at $1.24 per correct answer. The slide matters more than the sticker.

• Grok 4.5 landed #6 at 59% and was the slowest model on the board today, a 464-second typical answer. Cheap per correct answer, but you wait for it.

• Nemotron 3 Ultra kept the open-weights lead at 57%, a hair ahead of the debuting Opus 5 and comfortably clear of the rest of the open field.

• Kimi K3 held mid-pack at #9, 48%, one of the steadier budget buys at $0.30 per correct answer.

• Muse Spark 1.1 brought up the rear at 29% on another clipped slate, its sixth incomplete board in a row. The API still can’t finish what the model starts.

Why the board moves at all

Only 27 of 36 order calls between returning models repeated today (τ = 0.50), the loosest board in a week, and the churn is real rather than mechanical: Sol falling two, Gemini climbing four, Fable dropping three, GPT-5.5 rising two.

Our comparability controls read today’s slate as statistically as hard as yesterday’s, and no model tripped the drift check, so those swings are the models and their serving on the day, not an easier or harder exam.

A τ this low with the difficulty held flat is what a genuine reordering looks like.

Two models left the board today and one arrived, but the reshuffle among the nine that stayed is the part that counts.

Why not just trust the benchmarks you’ve seen?

Public benchmarks leak into the next training run, so scores inflate and start measuring memorization instead of ability.

They’re snapshots: true the day they’re taken, wronger every day after.

We generate fresh, novel problems every day. No model can have trained on a problem that didn’t exist until this morning, which is how a brand-new flagship can be measured against five-week incumbents the same day it ships, and land where the problems say it lands rather than where the launch says it should.

Subscribe, and each morning you’ll get the day’s best model, plus the fastest and cheapest way to get that performance, in your inbox before you start work.

Subscribe now

TL;DR

• Best closed model: GPT-5.6 Sol. 70%, a fourth straight crown at $0.22 per correct answer. But its lead over second place is now a coin-flip: 6-5 over Fable 5, significant only against the budget tail.

• Cheapest closed model: Grok 4.5, on the sticker. $0.20 per correct answer at 57.5%, though Sol matches it within two cents while solving a fifth more problems, so the crown is still the better buy.

• Best open-weights model: Nemotron 3 Ultra. 54%, holding the open lead for a second day despite provider-side errors again nicking its slate.

• Shuffled at the top. 46 of 55 order calls between returning models repeated (τ = 0.67): Fable climbed two seats back to the podium while Gemini gave back its whole debut and dropped three to sixth.

• Today’s tidbit: the podium collapsed. Sol, Fable, and GPT-5.5 finished 70 / 68 / 68, with Opus a step back at 62. The four best models on the board are separated by eight points, the tightest top we have measured.

What to do with today’s board

Sol stays the default, and now the safe value pick too: nothing outscores it, and nothing worth running is meaningfully cheaper per correct answer. The live question is what to do with Anthropic’s flagship. Fable 5 climbed ten points overnight into a genuine #2, but it still charges $1.16 per correct answer, roughly five times Sol’s rate, and it lost the head-to-head to Sol.

If you’re paying Fable or Opus prices for competitive-coding-style work, run Sol or GPT-5.5 in parallel this week and you’ll get equal-or-better results at a fraction of the bill. Don’t over-read yesterday’s Gemini hype; it regressed a full three seats today, so give it a third board before you rewire anything. And bench Qwen 3.7 Max until it can finish on time, because nearly a quarter of its attempts ran out the clock.

That’s our read. Argue with it; tomorrow’s board is the rebuttal.

isitjustme leaderboard, July 23, 2026

Three stories lead today’s board; the rest is quick hits.

1. A fourth crown, and the closest one yet

Sol holds #1 at 70%, its fourth morning in a row on top. But the margin that used to be a clean sheet is now a coin-flip. Against Fable it went 6-5, one problem, and against GPT-5.5 just 2-1.

The separation only reappears further down: 9-1 over Kimi (p=0.02), 10-1 over Qwen (p=0.01), 12-0 over GLM, 16-0 over Muse.

So the read is unchanged where it matters for your wallet. Sol still wins the budget half decisively and still books results at the cheapest rate of any model you’d actually deploy. What changed is the shape at the top: it’s no longer Sol and a gap but Sol and a cluster, three models inside two points, with the leader winning its closest matchup by a single solve.

Elo with confidence intervals, July 23, 2026

2. Fable storms back to the podium

The biggest mover on the board is Anthropic’s flagship. Fable 5 jumped from fourth to #2 at 68%, a ten-point climb overnight, and it earned it: 8-3 over Gemini, 6-4 over Opus, and a dead-even 5-5 with GPT-5.5 that it wins only on the Elo tiebreak.

Our day-to-day comparability controls read this morning’s slate as equal in difficulty to yesterday’s, so the leap is the model and its serving, not an easier board.

The catch is the one that has dogged Fable all month: it is still the priciest results on the table at $1.16 per correct answer, roughly five times Sol’s, and it still lost its one meeting with the champion. A stronger Fable is real news for anyone locked into the Anthropic stack, but it is not yet a reason to pay the premium over the model beating it.

Head-to-head record, July 23, 2026

3. Gemini’s debut hangover

Yesterday Gemini 3.6 Flash walked in at #3 and owned the stopwatch. Today it gave the whole debut back, dropping three seats to #6 at 55%, its head-to-heads with the top now underwater: 3-8 against Fable, 2-7 against GPT-5.5, 2-8 against Sol.

The one thing it kept is the speed, a typical 73-second answer that is still the fastest on the board, a hair ahead of Fable’s 75 and well under Sol’s 118.

This is the asterisk that debuts usually arrive with and this one skipped: a strong first impression is one board, and the second is where you find out whether it holds. Gemini’s second board didn’t. It’s still a fast, cheap, mid-table option worth a side-by-side, but the “above both Anthropic flagships” line from yesterday lasted exactly one day.

Quick hits

• Opus 4.8 ticked up a seat to #4 at 62%, but it sits behind both GPTs and its own sibling, and it’s the slowest of the leaders at a 302-second median. Roughly level on results with the pack ahead, and well behind on latency.

• Grok 4.5 owns the cheapest sticker on the board at $0.20 per correct answer, but at 57.5% and a 374-second median it’s a value play, not a default. You’re trading accuracy and time for the last two cents.

• Qwen 3.7 Max melted down: nearly a quarter of its attempts ran out the clock, dragging its median to a board-worst 720 seconds. When it answers it’s fine; too often it doesn’t answer in time.

• GLM-5.2 landed second-to-last at 40%, yet our comparability controls actually read its underlying level as higher than its baseline, not lower, so today’s score is the serving and the clock (three timeouts, three errors), not a capability drop.

• Kimi K3 climbed two seats to #8 at 49%, quietly one of the better budget buys at $0.25 per correct answer.

• Muse Spark 1.1 improved to 34% and again on an incomplete slate, its fifth clipped board in a row. The model keeps inching up while the API keeps standing still.

• Nemotron 3 Ultra held the open-weights lead at 54%, its recurring provider-side errors trimming the slate but not the crown.

Why the board moves at all

46 of 55 order calls between returning models repeated today (τ = 0.67), and the nine that flipped cluster at the top: Fable’s two-seat surge, Gemini’s three-seat slide, and an Opus-over-Grok swap account for nearly all of them.

That’s the useful reading of a middling τ. The movement isn’t noise smeared across the table; it’s concentrated in the handful of models that actually moved. The comparability controls tell us the slate was statistically as hard as yesterday’s, so Fable climbing, Gemini falling, and Sol holding are the models and their serving, not the problems.

A day like this, with a compressed top and one big climb against one big slide, is exactly why we re-measure every morning instead of trusting last week’s ranking.

Solve-rate heatmap, July 23, 2026

Why not just trust the benchmarks you’ve seen?

Public benchmarks leak into the next training run, so scores inflate and start measuring memorization instead of ability. They’re snapshots: true the day they’re taken, wronger every day after.

We generate fresh, novel problems every day. No model can have trained on a problem that didn’t exist until this morning, which is how a model that surged ten points overnight and one that surrendered a debut can both be caught on the same board, the same week, on equal footing.

(Today’s board costed $126.75 in compute.)

Subscribe, and each morning you’ll get the day’s best model, plus the fastest and cheapest way to get that performance, in your inbox before you start work.

Subscribe now

TL;DR

• Best closed model: GPT-5.6 Sol. 70%, a third straight crown at $0.22 per correct answer, with decisive wins over the entire budget half of the board.

• Cheapest closed model: Sol, by a whisker. Gemini 3.6 Flash debuted at $0.23 per correct answer, two tenths of a cent behind. The value race just became a race.

• Best open-weights model: Nemotron 3 Ultra. 50%, with a caveat: provider-side errors clipped a fifth of its slate.

• Steady board. 39 of 45 order calls between returning models repeated (τ = 0.73), and our comparability checks read today’s slate as equal in difficulty to yesterday’s, so the moves you see are real.

• Today’s tidbit: the debutant owns the stopwatch. Gemini’s typical answer took 73 seconds, the fastest we have ever measured, from a model that also finished top-three on accuracy.

  Kimi K3 continues to disappoint due to insufficient infra.
  
  Gemini 3.6 Flash was given the compute it needs.
  
  Don’t trust the headlines, look at the numbers below.

What to do with today’s board

Sol stays the default: three crowns, cheapest results, and nothing above it. The new decision is Gemini 3.6 Flash. On one day’s evidence it delivers near-Sol accuracy at Sol’s price in roughly half Sol’s time, so if your workload is latency-sensitive, start a side-by-side today; if it holds a second board we have a genuine two-horse race at the top of the value chart.

Fable’s premium remains hard to defend at $1.40 per correct answer against a debutant that beat it 7–5 while charging a sixth as much. And treat Muse Spark as benched until its API can finish a slate: that’s four clipped boards in a row.

That’s our read. Argue with it; tomorrow’s board is the rebuttal.

Two stories lead today’s board; the rest is quick hits.

1. Google shows up, and the stopwatch changes hands

First day on the board for Gemini 3.6 Flash, and it did three things at once.

It finished #3 at 62.5% on a full slate, above both Anthropic flagships, taking Fable 7–5 and Opus 6–3 on the problems that separated them.

It matched the champion’s economics at $0.23 per correct answer.

And it answered in a typical 73 seconds, the fastest figure we have ever recorded, comfortably ahead of Fable’s 105 and about half of Sol’s 136.

The gap to the GPTs is real (1–4 against Sol, 3–6 against GPT-5.5, neither significant yet), so this is a strong third place rather than a threat to the crown. But debuts usually arrive with an asterisk, and this one arrived with a full slate, a healthy drift check, and the best latency on the board.

We have not seen a cleaner first day.

2. Sol’s three-peat, with a memento mori

70%, a third consecutive #1, and the usual clean sheet against the budget field:

• 8–1 over Grok (p=0.04)

• 12–0 over Kimi K3

• 8–0 over Nemotron

• 10–0 over Qwen

• 16–0 over Muse

All decisive.

Against its sibling it went 3–1, and against the debutant 4–1.

The streak now equals the longest we have recorded, which is the cautionary part: Fable’s three-day run ended the morning after it looked inevitable, at the hands of a model nobody was watching. Today a model nobody here was watching just took third place with better latency and identical pricing.

Sol earned every crown; it should still sleep lightly.

Quick hits

• Fable 5 stopped the bleeding at 57.5%, unchanged from yesterday, still #4 and still the priciest results at $1.40 per correct answer. Flat is better than falling. It is not a case for the premium.

• Opus 4.8 slipped to #5 at 55%, statistically level with Fable (4–3). The two Anthropic flagships are now interchangeable on results and separated only by price.

• GLM-5.2 gave back some of yesterday’s high at 42.5%, though our comparability checks still read its underlying level as elevated; nine of its problems ran out the clock today, which is the real story.

• Nemotron 3 Ultra took the open-weights lead at 50% on 32 scored problems; the other eight died to provider-side errors, its recurring weakness.

• Muse Spark 1.1 improved again to 30%, and again on a clipped slate. Four boards, four incomplete slates. The model is inching up; the API is standing still.

Why the board moves at all

39 of 45 order calls between returning models repeated today (τ = 0.73), and the swaps concentrate below the top four, where a one-seat mechanical shift from the debutant’s arrival mixes with GLM and Kimi K3 easing back.

The more interesting reading is the one our comparability checks gave us: today’s slate was statistically equivalent in difficulty to yesterday’s, which means the score movements you see, Sol holding, Fable flattening, GLM retreating, are the models and their serving, not the problems.

Days like this are why we run the same measurement every morning instead of trusting last week’s.

Subscribe now

Why not just trust the benchmarks you’ve seen?

Public benchmarks leak into the next training run, so scores inflate and start measuring memorization instead of ability. They’re snapshots: true the day they’re taken, wronger every day after.

We generate fresh, novel problems every day. No model can have trained on a problem that didn’t exist until this morning, which is why a model that launched into public availability can be measured against five-week incumbents on equal footing the same week.

(Today’s board took about $111 in compute.)

Subscribe, and each morning you’ll get the day’s best model. Plus the fastest and cheapest way to get that performance. In your inbox before you start work.

Subscribe now

TL;DR

• Best closed model: GPT-5.6 Sol. 72.5%, a second straight crown, and a second straight day as the cheapest result on the board at $0.22 per correct answer.

• Cheapest closed model: also Sol, again. Grok 4.5 is a few cents behind ($0.26) at 22 points fewer. Two boards running, the crown and the value column share a row.

• Best open-weights model: GLM-5.2. A first. 45%, its best board ever, #6 overall and three seats up.

• Steady board. 40 of 45 order calls between returning models repeated (τ = 0.78). The one real mover was GLM climbing past Qwen, Kimi K3, and Nemotron.

• Today’s tidbit: across nine opponents and every problem where results diverged, Sol lost a total of three problems all day. One to Fable, two to GLM, none to anyone else.

What to do with today’s board

Sol is the default again, and the argument got stronger: champion, cheapest result on the board, and it dropped three discordant problems all day.

If you’re on GPT-5.5 the switch remains free.

Fable needs a reason now, and “fastest” stopped being one today; if you’re paying its premium, run Opus beside it this week and check whether you notice the difference anywhere but the invoice.

Open-weights buyers should actually look at GLM-5.2 for the first time, with the caveat that its history is volatile and one great board is one great board.

And Qwen users with deadlines: eleven problems ran out the clock today. That risk is yours too.

That’s our read. Argue with it; tomorrow’s board is the rebuttal.

Subscribe, and each day you’ll get the day’s best model, plus the fastest and cheapest way to get that performance, in your inbox.

Subscribe now

Two stories lead today’s board; the rest is quick hits.

1. Sol’s second crown was more convincing than its first

Yesterday Sol took #1 on a tiebreak. Today it took #1 on the merits: 72.5%, with statistically decisive wins over six of its nine opponents.

9–0 over Grok (p=0.004), 14–0 over both Kimi K3 and Nemotron (p=0.0001), 12–0 over Qwen, 18–0 over Muse, 13–2 over GLM.

Against the models above the budget line it went 2–0 on GPT-5.5, 4–0 on Opus, and 7–1 on Fable (p=0.07, a hair short of significance).

Add the price, $0.22 per correct answer for the second day straight, and the case writes itself.

The one caution we’d offer: two crowns in a row is a streak, and we watched a three-day Fable streak end badly last week. Streaks are rented.

2. Fable’s slide reaches four boards

The sequence now reads 80%, 70%, 60%, 57.5%, and the rank line reads #1, #2, #4, #4.

Today Fable took just 1 of 8 discordant problems against Sol, at $1.69 per correct answer, its worst value figure yet and nearly eight times the champion’s.

Even the consolation title moved: for the first time in six boards, the fastest typical answer belonged to someone else (Muse Spark, on its reduced slate, with Fable close behind).

We keep looking for the floor here. A 57% day from the priciest model on the board is not it, and the only argument left for the premium is a speed edge measured in seconds.

Anthropic loyalists should be watching Opus, which quietly outrank Fable for a second straight day at a third less per answer.

Quick hits

• GLM-5.2 finished #6 at 45%, its best board since we started tracking it, and now leads the open-weights field outright. Five boards ago it solved one problem. The serving change behind it keeps paying.

• GPT-5.5 held #2 at 67.5% with a significant 8–1 over Grok. Its $0.43 per correct answer is double Sol’s, a sentence nobody expected to type two weeks ago.

• Grok 4.5 fell to 50%, went 0–9 against Sol on their discordant problems, and its frozen-set drift score has sagged for a third board. Something may have changed behind that name too.

• Qwen 3.7 Max ran out the clock on eleven problems, its worst count yet. Its slow tail is now the main thing standing between it and the top half.

• Muse Spark 1.1 improved to 28% on a slate again clipped by provider-side errors (about a fifth of its problems went unscored). Fast, cheap, last, and slightly less last than before.

Why the board moves at all

40 of 45 order calls between returning models repeated today (τ = 0.78), which is the steady end of our range, and the movement that did happen was concentrated in one climber: GLM stepping past Qwen, Kimi K3, and Nemotron in a single day.

The top four held their seats exactly.

When a board this stable still shows a 25-point swing at one model (Grok’s 9-point drop, GLM’s 5-point rise), the explanation usually lives in serving stacks rather than model weights, and our frozen-problem comparability checks pointed the same direction today.

Fresh problems daily is what lets us say that with a straight face.

Why not just trust the benchmarks you’ve seen?

Public benchmarks leak into the next training run, so scores inflate and start measuring memorization instead of ability.

They’re snapshots: true the day they’re taken, wronger every day after.

We generate fresh, novel problems every day. No model can have trained on a problem that didn’t exist until this morning, which is how a two-week-old flagship defends a crown and a written-off budget model climbs into the top six on the same page.

(Today’s board took about $123 in compute.)

Subscribe now

TL;DR

• Best and cheapest closed model: GPT-5.6 Sol. 72.5%, #1 on a tiebreak over its own sibling GPT-5.5 (identical solve rate, 3–3 head to head), at $0.22 per correct answer. The champion is also the cheapest result on the board.

• Best open-weights model: Nemotron 3 Ultra. 54% at $0.35 per correct answer, ahead of Kimi K3 (45%) in its debut week.

• Shakeup. 36 of 45 order calls between returning models repeated since Thursday’s board (τ = 0.60). The movement was real: the GPTs went one-two, Opus climbed to #3, Fable slid to #4, Qwen fell three seats.

• Today’s tidbit: Muse Spark 1.1 returned, and this time it answered. Most of its slate scored (a handful of problems still lost to provider-side errors), fast (~112s typical) and cheap ($0.37 per answer). It also finished last at 23.5%, swept 20–0 by the champion.

Share

Three stories lead today’s board; the rest is quick hits.

1. Sol’s arc completes: quiet-quitter to champion at the best price we’ve printed at #1

A week ago Sol lost to its own predecessor 0–4 and we called it a speed tax. Thursday it posted the board’s best economics and we said we’d wait for a second day.

The second day came in at #1, 72.5%, dead even with GPT-5.5 at 3–3 and ahead only on strength of schedule.

For your bill, the number is $0.22 per correct answer: the cheapest results on the entire board, from the model wearing the crown.

Against the rest of the field it wasn’t subtle. 8–0 over Qwen (p=0.008), 11–0 over Kimi K3 (p=0.001), 20–0 over Muse, 7–1 over Grok.

The GPT lane now holds first and second place, and choosing between the siblings has become a matter of taste: identical results, with Sol a shade faster and cheaper today. A week ago that sentence would have been a joke.

2. Fable’s premium is now a speed premium only

Fable’s last three boards read 80%, 70%, 60%, and its rank went #1, #2, #4.

Today it lost 3–8 against both GPTs on the problems that separated them (p=0.23, so not yet significant, but pointing one direction), fell out of the top three entirely, and remained the priciest model on the board at $1.15 per correct answer, five times Sol’s rate for twelve points fewer.

What’s left of the case is speed, and that part is real: fastest again at ~102s typical, the fifth straight board it has held that title.

If your workload pays for the fastest excellent answer, Fable still cashes. If it pays for the best answer, this week the board says that’s a GPT, at a fifth the price.

3. Muse Spark came back, and now we know what it is

Ten days after a launch where roughly 9 in 10 of its requests failed, Muse Spark 1.1 rejoined the roster, and the infrastructure mostly behaved this time: the bulk of its slate produced scoreable answers, with a handful of problems lost to provider-side errors.

The results are another matter. Last place at 23.5%, swept 20–0 by Sol, 19–1 by GPT-5.5 and Opus, 16–0 by Fable, each of those statistically decisive.

The odd part is the shape: it answers in about 112 seconds, second fastest on the board, at $0.37 per correct answer. It gets the hard ones wrong quickly and cheaply.

Not a contender on this board, but Meta is presumably iterating on it, and we’ll keep scoring it.

Quick hits

• Opus 4.8 climbed to #3 at 67%, its best rank in over a week, including a significant 9–1 over Kimi K3. Its results stay the second priciest at $0.88 per answer.

• Grok 4.5 held #5 at 57% and matched Sol’s per-answer price ($0.22). Sol’s 7–1 edge on their discordant problems says the same dollar buys a different ceiling.

• Qwen 3.7 Max gave back half of Thursday’s surge, 70% to 52%. The slowest answers on the board finally cost it: a quarter of its slate ran out the clock and scored as failures.

• Kimi K3 beat its sibling for the first time, 45% to K2.7’s 40%, on full slates. Both trail Nemotron, and K3’s answers stay slow.

• GLM-5.2 held its new level at 40%, two boards running. The serving change was real.

Why the board moves at all

36 of 45 order calls between returning models repeated since Thursday’s board (τ = 0.60), and the exceptions were concentrated moves rather than noise: the GPTs rising together, Fable stepping down two, Qwen’s surge partly unwinding, the Kimi siblings swapping.

Serving stacks keep shifting under several of these names (Thursday’s thirty-point jumps, today’s partial retreats).

A board that re-measures everything on brand-new problems each morning is the only way to see what these names deliver today, rather than what they delivered whenever someone last ran an eval.

Why not just trust the benchmarks you’ve seen?

Public benchmarks leak into the next training run, so scores inflate and start measuring memorization instead of ability. They’re snapshots: true the day they’re taken, wronger every day after.

We generate fresh, novel problems every day. No model can have trained on a problem that didn’t exist until this morning, which is how a model can go from quiet quitting to champion inside a week, and how a relaunched model gets judged on what it does now rather than on its press cycle.

What to do with today’s board

The GPT lane is the default this morning.

Sol at $0.22 per correct answer is the champion and the cheapest result at once.

If you’re on GPT-5.5, switching costs nothing and gains a little speed. If you’re on anything pricier, the burden of proof just moved to your invoice.

Fable is a specialist buy now: pay the premium if the fastest excellent answer is the product.

Kimi’s infrastructure limitations are potentially holding back amazing performance. Let them sort it out. If you need to build around an open weights model, try Nemotron or

Skip Muse for real work, since fast and cheap doesn’t survive 20–0, but keep an eye on it; Meta ships fast.

And if your Qwen workloads have deadlines, that 52% understates the risk, because its slow tail is now failing problems outright.

That’s our read. Argue with it; tomorrow’s board is the rebuttal.

TL;DR

• Best closed model: GPT-5.5 at 75%. It takes its second #1 of the week and, for once, the champion is also a value pick at $0.38 per correct answer.

• Cheapest closed model: GPT-5.6 Sol at $0.22 per correct answer and a 70% solve rate, the best economics on the entire board today.

• Best open-weights model: Nemotron 3 Ultra at 48%, finally measured on a full slate, at $0.37 per correct answer.

• The shakeup: only 29 of 36 order calls between returning models repeated, for τ = 0.61. Nearly every meaningful swap happened in the bottom half.

• Today’s tidbit: Qwen 3.7 Max jumped 37 percentage points into a tie for the second-best solve rate. GLM-5.2 went from one solve to 38%. Our frozen-problem controls indicate that the problems did not get easier. The serving paths changed.

Two stories lead today’s board. The rest is quick hits.

1. GPT-5.5 is champion and value pick at the same time

Fable’s streak ends at three.

GPT-5.5 takes #1 at 75%, giving it a weekly run of:

#1 → #4 → #2 → #1

For the first time this week, the crown and the value title sit in the same row. GPT-5.5 cost $0.38 per correct answer, roughly one-third of what Fable charged while producing more correct answers.

The margin still deserves honesty. The top four are a statistical pileup:

• 3–1 over Fable

• 3–1 over GPT-5.6 Sol

• 3–1 over Qwen 3.7 Max

Those are coin flips, not decisive wins. The flagships agreed on nearly every problem today.

Fable still owns speed, with a typical response time of roughly 102 seconds, making it the fastest model again. Three days ago, the premium argument was:

Best, fastest, and statistically significant.

Today, it is simply:

Fastest.

2. The bottom half is unrecognizable

The largest moves did not happen among the flagships.

They happened here:

• Qwen 3.7 Max: 32% → 70%

• GLM-5.2: one solve → 38%

• Nemotron 3 Ultra: 48% on a full slate

Models do not learn this much overnight.

The result can be materially different capability under the same label.

That cuts both ways for consumers.

The Qwen available today is a genuinely different proposition from the one that scored 32% yesterday. At the same time, any evaluation performed on last week’s serving stack may already be stale.

Quick hits

GPT-5.6 Sol quietly posted the board’s best economics

Sol reached 70% at just $0.22 per correct answer, while operating at nearly Fable-level speed.

That is the best price-performance result on today’s board. If it holds for a second day, the conversation around what users are actually paying for starts over.

Grok 4.5 slowed down sharply

Grok held fifth place at 68%, but its typical response stretched to roughly 10 minutes, more than triple its usual pace.

Latency-sensitive users should check current performance before placing large workloads in the queue.

Opus 4.8 remains directionless

Opus slipped to sixth at 62%, falling below Qwen while remaining near the top of the cost table at $1.09 per correct answer.

Its week has not been disastrous. It has simply lacked a clear argument.

Kimi K3’s debut is now complete

Kimi K3 finished at 40% on a full slate.

The diagnosis changed once the full run completed. Its main problem is speed, not reliability. Nearly 1/4^th of its slate ran out the clock (1,800s) and counted as failures.

On problems both Kimi models completed, K2.7 edged K3 by 6–5.

A faster K3 could become a serious contender. This version is not there yet.

Kimi K2.7 Code loses the open-weights crown

Kimi K2.7 finished at 42%, losing the open-weights lead to Nemotron.

Its successor, Kimi K3, finished just one solved problem behind it.

Why the board moved

This was a genuine shakeup day.

Only 29 of 36 order calls between returning models repeated, producing τ = 0.61, the lowest stability reading in a week.

But the location of the movement matters.

The top five mostly reshuffled within statistical noise. The real volatility appeared in the bottom half, precisely where the serving-side changes landed.

That is the board doing its job.

When the system actually being served under a model’s name changes, yesterday’s score stops being strong evidence. Only a fresh measurement, using problems nobody could have trained on, reveals what users are buying today.

Transparency note

Today’s board required two measurement passes:

• First pass: roughly 8.8 hours

• Second pass: roughly 7.0 hours

• Total wall-clock measurement time: approximately 15¾ hours

The second pass completed Kimi K3’s full slate after provider-side API errors erased roughly one-third of its debut run. It also re-attempted a small number of problems that encountered infrastructure failures for other models.

Every model shown above is scored on its complete slate.

One verdict changed after re-examination: a Nemotron pass did not reproduce, moving its score from 50% to 48%.

Everything else held.

Why not trust the benchmarks you have already seen?

Public benchmarks leak into future training runs. Scores rise, but the metric gradually shifts from measuring capability to measuring familiarity.

They are also snapshots. Even when a benchmark is valid on the day it is published, it becomes less representative as models and serving systems change.

Today’s board shows that model drift is only half the problem.

The serving stack behind a model name can change overnight and move observed capability by more than 30 percentage points.

We generate fresh, novel problems every day and re-measure every model each morning.

Today’s board cost approximately $121 in compute. But you can subscribe for free!

Subscribe now

What to do with today’s board

The GPT lane won today’s argument.

GPT-5.5 is the champion at 75% and a value pick at $0.38 per correct answer. If you already use it, there is little reason to move. If you are paying more elsewhere, today is a reasonable day to ask what the premium is buying.

Watch GPT-5.6 Sol closely. $0.22 per correct answer at 70% is the best economics we have printed, pending confirmation on a second day.

Fable users are not wrong. It remains elite and continues to be the fastest model on the board. But the winning streak is over, the statistically significant separation is gone, and the premium now needs another justification.

Most importantly, treat every budget model opinion formed before today as expired.

The Qwen, GLM, and Nemotron being served now are not the same systems users evaluated yesterday.

That is our read.

Argue with it in the comments below. Tomorrow’s board is the rebuttal.

TL;DR

• Best closed model: Fable 5, #1 for a third straight day at 80%. It did not lose a single problem to anyone ranked sixth or below: 31–0 over GLM, the most lopsided result we have ever published, plus 19–0 over Qwen, 18–0 over Nemotron, and 12–0 over Kimi.

• Cheapest closed model: Grok 4.5, $0.37 per correct answer at 58%. The value remains strong, but it has now taken three straight statistically significant losses to Fable, including 3–12 today (p=0.04).

• Best open-weights model: Kimi K2.7 Code, 49% at $0.15 per correct answer, unchallenged again.

• A steady board: 33 of 36 order calls between returning models repeated (τ = 0.83). The only swaps were GPT-5.5 climbing past Sol and Opus, and Nemotron moving past Qwen.

• Today’s tidbit: GLM-5.2 solved one problem all day, its second one-solve day this month. Its recent oscillation, 2% → 22% → 12% → 2%, is the strangest signature we track.

Two stories lead today’s board. The rest is quick hits.

1. Day three: Fable’s floor is perfect, but its ceiling has company

The streak held.

Fable finished at 80%, ranked #1 for a third straight day, and once again produced the fastest typical answer on the board at roughly 107 seconds.

Against the bottom half of the field, it was flawless: four budget models, 40 separating problems, zero losses.

Its 31–0 result against GLM-5.2 is the most lopsided head-to-head we have ever published.

But the honest story at the top is less decisive.

Yesterday’s daylight narrowed considerably:

• 5–3 over GPT-5.5, down from 8–1 yesterday (p=0.73)

• 5–2 over GPT-5.6 Sol

• 7–2 over Opus 4.8

• 12–3 over Grok 4.5 (p=0.04)

Only the Grok matchup remained statistically significant, and this was Fable’s third significant win over Grok in a row.

Three days in, Fable’s claim to #1 rests on two things: a sustained winning streak and a perfect floor against weaker models. It does not yet have clear statistical separation from the other flagships.

The price also moved in the wrong direction. At $1.53 per correct answer, this was Fable’s most expensive day yet, widening the value gap between the champion and the models immediately below it.

2. The GPT seesaw swings back: GPT-5.5 jumps from #4 to #2

Monday’s champion. Yesterday’s #4. Today’s #2.

GPT-5.5 rebounded to 75%, passed both Sol and Opus, and edged its successor 3–2 on their discordant problems.

It also delivered the runaway best value among the top four: $0.76 per correct answer, roughly half the cost of the models ranked immediately around it.

Sol did not slump. It improved slightly from yesterday’s 70% to roughly 72% today. GPT-5.5 simply passed it again.

The last three boards now look like this:

• GPT-5.5: #1 → #4 → #2

• GPT-5.6 Sol: #5 → #2 → #3

Each individual swing has had a plausible explanation. Taken together, though, the honest conclusion is simpler: the two models remain statistically interchangeable on results and durably different on price.

Until one of them produces a sustained run of separation, treat the GPT rank churn as weather.

Quick hits

Opus 4.8: displaced, not degraded

Opus slipped one seat to #4 at 68%, but it did not meaningfully get worse. The GPT models moved around it.

It remained statistically level with both, going 2–4 and 3–6 on separating problems, but at $1.21 per correct answer, it was never the value pick.

Grok 4.5: still cheap, no longer a budget flagship

Grok held #5 at 58% and retained the best closed-model price at $0.37 per correct answer.

But three consecutive statistically significant losses to Fable settle the “budget flagship” question for now.

It is budget. It is not currently flagship.

Nemotron 3 Ultra: same accuracy, five times slower

Nemotron retained the cheapest-results crown at $0.04 per correct answer, with every cell successfully judged for once.

The problem was latency.

Its typical answer took 12.6 minutes, compared with 2.5 minutes yesterday, a slowdown of nearly fivefold at roughly the same accuracy.

Something appears to have changed server-side. Watch this one.

Qwen 3.7 Max: the climb ends

Qwen’s three-board rise stalled. Its solve rate fell from 38% to 32%, allowing Nemotron to pass it again.

GLM-5.2: one solve

One correct answer all day.

There is no price at which a 2% solve rate becomes a deal.

Why the board moves at all

This was another steady day where it mattered.

33 of 36 order calls between returning models repeated, producing a Kendall τ of 0.83.

All three swaps came from two movers:

1. GPT-5.5 rebounded past Sol and Opus.

2. Nemotron edged past a fading Qwen.

The top seat remained unchanged for a third straight day.

Meanwhile, the largest absolute score movement, GLM’s 20-point collapse to one solve, happened in the part of the board where it changes no real decision.

Fresh problems cause raw scores to breathe. Three consecutive days with the same model on top is what signal begins to look like.

Why not just trust the benchmarks you have already seen?

Public benchmarks eventually leak into training data.

As that happens, scores inflate and begin measuring memorization rather than capability. A benchmark may be accurate when it launches, then become less informative with every subsequent training run.

We generate fresh, novel problems every day.

No model can have trained on a problem that did not exist until that morning. That lets us measure a three-day winning streak, a two-day crown, and a one-solve collapse on exactly the same footing.

Today’s board required approximately $158 in compute.

Subscribe now

What to do with today’s board

Nothing at the top changed enough to force a switch.

Already using Fable? Three straight days at #1 says stay.

Already using GPT-5.5? Today’s #2 finish at $0.76 per correct answer, roughly half the price of the models around it, says you are still in the right seat. Sol’s value case weakened today rather than strengthening.

The one real operational warning is Nemotron. If you chose it partly for speed, check your production latencies. It ran nearly five times slower today without delivering better accuracy.

And if any of your workloads still route to GLM-5.2, today it answered one question correctly.

That is our read. Argue with it. Tomorrow’s board is the rebuttal.

TL;DR

• Best closed model: Fable 5 — 82%, with a second straight day of real separation: 8–1 over GPT-5.5 (p=0.04), 12–2 over Grok 4.5 (p=0.01), and clean sweeps across the budget half of the board, including 28–0 over GLM-5.2.

• Cheapest closed model: Grok 4.5 — $0.34 per correct answer at 57%. The value case holds, but Fable has now beaten it decisively on two consecutive boards.

• Best open-weights model: Kimi K2.7 Code — undisputed today at 49% and $0.13 per correct answer, while Nemotron fell eighteen points.

• A steady board. 33 of 36 order calls between returning models repeated (τ = 0.83). Nearly every swap came from two movements: Sol and Opus stepping over GPT-5.5, and Nemotron falling beneath Kimi and Qwen.

• Today’s tidbit: GPT-5.5 went from #1 on Monday to #2 yesterday to #4 today. The board giveth.

Two stories lead today’s board. The rest is quick hits.

1. Fable 5, day two: the gap is starting to look structural

One day of daylight can be a good draw. Two starts to look like a pattern.

Fable 5 followed yesterday’s breakout with an 82% solve rate, and the statistical picture moved from “ahead” to “significantly ahead of almost everyone.”

It beat:

• GPT-5.5 8–1 (p=0.04)

• Grok 4.5 12–2 (p=0.01)

• Kimi K2.7 Code 14–1

• Qwen 3.7 Max 19–1

• GLM-5.2 28–0

That last result is the most lopsided head-to-head we have ever printed.

Only Sol (7–2, p=0.18) and Opus (7–1, p=0.07) escaped statistical significance, and neither by much.

Fable also remained the fastest model on the board, producing a typical answer in roughly 106 seconds. At $1.26 per correct answer, it even cost less per successful result than Sol one seat below it.

Two days ago, paying for Fable became defensible. Today, it is starting to look like the default.

2. The GPT seesaw: Sol completes the redemption arc as 5.5 gives back the crown

On Monday, Sol lost 0–4 to its own predecessor. We said it needed to prove the premium for two consecutive days.

Consider it proven.

Sol finished #3 yesterday and #2 today, reaching a 70% solve rate. For the first time, it also finished above GPT-5.5, edging their direct matchup 5–3.

Meanwhile, GPT-5.5 went from champion to runner-up to #4 in 48 hours, shedding roughly seven and a half percentage points along the way.

Before migrating anything, however, the economics still favor 5.5.

GPT-5.5 delivered its 65% solve rate at $0.87 per correct answer, compared with $1.38 for Sol. That makes 5.5 the cheapest source of correct answers in the top four by a wide margin.

Their direct matchup also remains statistically indistinguishable from a coin flip (p=0.73).

Sol has earned the flagship label. GPT-5.5 still owns the value label.

Quick hits

• Opus 4.8 quietly climbed to #3 at 68%, finishing effectively level with Sol in their head-to-head (3–4) and GPT-5.5 (4–3). Steady and capable, but still not the value pick at $1.21 per correct answer.

• Grok 4.5 held #5 at 57% and retained the best closed-model price on the board at $0.34 per correct answer. But it has now suffered two consecutive significant losses to Fable. The “level with the leaders” era is over. The “80% cheaper than them” era continues.

• Nemotron 3 Ultra surrendered its spike, falling from 54% to 36% and sliding below both Kimi and Qwen. It also lost cells to provider-side API errors for the fourth consecutive board, making it the least reliable model we track.

• Qwen 3.7 Max improved for a third straight board: 33% → 35% → 38%. It remains tied with GLM as the slowest model on the board.

• GLM-5.2 posted its worst result yet at 12%, including a 28–0 sweep by Fable.

Why the board moves at all

This was a steady day.

33 of 36 order calls between returning models repeated, producing a Kendall rank correlation of τ = 0.83.

The three changes all came from two genuine movements:

1. GPT-5.5 eased beneath Sol and Opus.

2. Nemotron’s Monday spike unwound, dropping it below Kimi and Qwen.

Nothing else moved.

The same fresh-problems-daily machinery that produced last week’s shakeup produced near-stasis today. That is exactly what the system should do when the models themselves are the thing that did not meaningfully change.

Why not just trust the benchmarks you’ve already seen?

Public benchmarks eventually leak into training data. Scores rise, and the benchmark gradually starts measuring memorization instead of ability.

They are also snapshots: accurate on the day they are collected, and less representative every day afterward.

We generate fresh, novel problems every day. No model can have trained on a problem that did not exist until that morning.

That is how a five-day-old flagship’s redemption arc, a champion’s rapid fall, and a two-day winning streak can all be measured on identical footing.

Today’s board required roughly $151 in compute.

Subscribe now

What to do with today’s board

If you moved to Fable after Monday’s board, stay there. Two consecutive days of statistically meaningful separation, combined with the fastest response time on the board, is about as strong a signal as this leaderboard can produce.

If you are choosing between the GPTs, Sol now appears to be the stronger model. But GPT-5.5 at $0.87 per correct answer remains the best value in the top four. Pay for Sol only when its roughly one-minute speed advantage is worth paying about 60% more per successful result.

Keep Grok for cost-sensitive workloads, but stop expecting it to tie the leaders.

For open weights, the answer today is simple: Kimi. Nemotron’s spike did not survive contact with a second day.

That is our read. Argue with it. Tomorrow’s board is the rebuttal.

TL;DR

• Best closed model: Fable 5. It finished at 80%, taking #1 with the first statistically significant daylight any leader has shown in a week. Fable went 10–2 against Grok 4.5 (p=0.04), 14–0 against Kimi (p=0.0001), and beat every flagship head-to-head.

• Cheapest closed model: Grok 4.5. It delivered correct answers for $0.34 each at a 60% solve rate. But read the quick hits: for the first time, a leader put Grok away decisively.

• Best open-weights model: disputed. Nemotron 3 Ultra ranks higher at 54%, but its slate was clipped by provider-side errors. Full-slate Kimi K2.7 Code finished at 45%, cost $0.14 per correct answer, and won their head-to-head 5–2. Our pick is Kimi.

• A steady board. 32 of 36 order calls between returning models repeated (τ = 0.78), the steadiest day in nearly a week. The swaps were Fable over GPT-5.5, Sol and Opus over Grok, and Nemotron trading places with Kimi.

• Today’s tidbit: the best model was also the fastest. Fable’s typical answer landed in about one minute and forty seconds, quicker than every budget model on the board. Accuracy and speed have never shared the same row before.

Two stories lead today’s board. The rest is quick hits.

1. Fable 5 didn’t just take the crown back. It broke the gridlock.

Three straight boards ended in coin flips at the top.

Today didn’t.

Fable 5 posted an 80% solve rate, opened 4.5 points of Elo daylight, and created the first statistically significant separation from the value pack since the top of the board seized up:

• 10–2 over Grok 4.5 (p=0.04)

• 14–0 over Kimi K2.7 Code (p=0.0001)

• 9–0 over Nemotron 3 Ultra (p=0.004)

Against the other flagships, the margins were friendlier but consistent:

• 6–3 over GPT-5.5

• 7–3 over GPT-5.6 Sol

• 7–3 over Opus 4.8

Fable beat every model on the problems that separated them.

It also did all of this while being the fastest model on the entire board. Its typical answer arrived in about 100 seconds.

At $1.15 per correct answer, Fable was also cheaper per successful result than Sol ($1.43) and Opus ($1.18).

The priciest sticker bought the best results, the fastest answers, and a mid-flagship realized cost.

Days like this are why the sticker-price column is not the cost column.

2. Sol answers yesterday’s embarrassment, halfway

Yesterday, GPT-5.6 Sol lost to its own predecessor 0–4.

We said: prove the premium, or the board keeps saying “buy 5.5.”

Today, Sol rebounded from #5 to #3 at 70%, and its head-to-head against GPT-5.5 evened out at 3–4 (p=1.0).

That completes half the assignment.

The other half remains open:

• GPT-5.5 still ranks higher.

• GPT-5.5 costs $0.80 per correct answer, compared with Sol’s $1.43.

• Sol produced the most expensive correct answers on today’s board.

• GPT-5.5 has now spent two straight boards above its successor.

One more day like today and Sol starts looking like a real flagship.

One more day of “statistically even, nearly twice the price,” and it is just a speed tax.

Quick hits

Grok 4.5 hits its first wall

Grok dropped to #5 at 60%, and Fable beat it 10–2, the first statistically significant loss Grok has taken since joining the board.

The value case survives. Grok still costs just $0.34 per correct answer and remains statistically level with GPT-5.5.

But “cheap and tied with the leaders” has now become just “cheap.”

Opus 4.8 posts a personal best

Opus recorded its best score on our board at 70% and finished dead even with Sol in their head-to-head, 4–4.

It was also the only model to leave problems unanswered, for the second board in a row.

Nemotron ranks above Kimi, with an asterisk

Nemotron 3 Ultra ranked #6 at 54%, but provider-side API errors clipped its slate for the third consecutive board.

It is the flakiest model we track.

Full-slate Kimi won their head-to-head 5–2. Nemotron still produces the cheapest correct answers anywhere, at $0.05 each, when its API shows up.

GLM-5.2 gives the gains back

GLM’s three-board climb reversed.

It finished at 18%, in last place, and was the slowest model on the board.

The bounce was noise.

Why the board moves at all

This was the steadiest board in nearly a week.

32 of 36 order calls between returning models repeated, producing a Kendall rank correlation of τ = 0.78.

The four swaps all happened at the seams:

• Fable edged past GPT-5.5 at the top.

• Sol and Opus stepped over a slipping Grok.

• Nemotron and Kimi traded places in the budget tier.

Individual scores still moved. Sol gained 12 percentage points, while GLM lost five.

That is exactly the pattern our day-to-day comparability controls are designed to referee.

Fresh problems change the difficulty costume every morning. A mostly stable ranking, with a few genuine movers, is what real signal looks like.

Why not just trust the benchmarks you’ve already seen?

Public benchmarks leak into subsequent training runs.

Scores inflate, and the benchmark gradually starts measuring memorization instead of ability.

They are also snapshots. They may be accurate on the day they are run, but they become less representative every day afterward.

We generate fresh, novel problems every morning. No model can have trained on a problem that did not exist until today.

That is how a four-day-old flagship’s redemption arc and a week-old champion’s counterpunch can be measured on identical footing, morning after morning.

What to do with today’s board

Today is the first board in a week where paying top dollar is clearly defensible.

Fable 5 was:

• The best model

• The fastest model

• Mid-pack on realized cost

If your work justifies flagship pricing at all, today it justifies Fable.

If it doesn’t, GPT-5.5 remains the sane default. It ranks #2, costs $0.80 per correct answer, and still has not lost to its more expensive successor.

Pause the Grok side-by-side if you started one based on our earlier results. A statistically significant 10–2 loss to the champion is the first real evidence against it. One more board like this ends the experiment.

For open weights, stay on Kimi. Do not move to Nemotron until it completes a full slate.

That’s our read.

Argue with it. Tomorrow’s board is the rebuttal.

Subscribe, and each morning you’ll get the day’s best model, plus the fastest and cheapest way to get that performance, in your inbox before you start work.

Subscribe now

TL;DR

• Best closed model: GPT-5.5. It finished #1 at 68%, dead even with Fable 5 on solve rate and head-to-head results (4–4). GPT-5.5 cost $0.86 per correct answer, compared with Fable’s $1.43.

• Cheapest closed model: Grok 4.5. It delivered a 62% solve rate for $0.32 per correct answer and remained statistically level with both leaders. It went 5–7 against Fable 5 (p=0.77).

• Best open-weights model: Kimi K2.7 Code. It scored 46% at $0.17 per correct answer and beat Nemotron 3 Ultra 4–1 on the problems where they disagreed.

• Today’s shakeup: 29 of 36 order calls between returning models repeated (τ = 0.61). GPT-5.5 climbed past Fable, Grok, and Sol; Opus 4.8 rebounded past Kimi; Nemotron and Qwen swapped places.

• Today’s tidbit: GPT-5.6 Sol, statistically level with the leaders on its Thursday debut, fell to #5 and went 0–4 against its own predecessor at nearly twice the realized cost. So far, the new-model premium is just a premium.

Two stories lead today’s board. The rest are quick hits.

1. GPT-5.5 and Fable 5 tied for the crown, but only one charges 40% less

Identical solve rates at 68%. A perfect 4–4 split on the problems that separated them. Elo ratings just three points apart.

This is about as tied as two models can get.

The ranking algorithm gave GPT-5.5 the top spot on strength of schedule, by the thinnest margin we have published. The bill was not nearly as close: GPT-5.5 delivered its half of the tie for $0.86 per correct answer, compared with $1.43 for Fable 5.

One seat below them, Grok 4.5 remains statistically level with both leaders. It scored 62%, went 5–7 against Fable (p=0.77), and cost just $0.32 per correct answer.

The podium is therefore less a performance ranking than a three-model argument about how much you enjoy paying.

The Elo chart below captures it well: one overlapping blur at the summit.

2. Sol’s crown lasted one news cycle

On Thursday’s debut, we wrote: “Watch this one. Debuts move.”

They move in both directions.

GPT-5.6 Sol finished #5 at 58% today, behind the predecessor it is supposed to replace. It went 0–4 against GPT-5.5 on the problems where they disagreed (p=0.13) and 3–7 against Fable 5.

Sol kept its speed advantage, producing a typical answer in about two minutes, although even Fable was faster today. Its verbosity, however, kept the meter running.

Sol cost $1.68 per correct answer, effectively tying Fable for the most expensive result on the board. That came from a model that just lost head-to-head against a sibling costing roughly half as much.

One day is not a verdict. But the question has flipped since launch:

Can Sol prove the premium, or will the board keep saying “buy 5.5”?

Quick hits

• Opus 4.8 rebounded hard: 38% → 60%, climbing three seats to #4. Its value still trails its performance, at $1.55 per correct answer, more than either model ranked above it.

• Nemotron 3 Ultra recovered: 22% → 40%, gaining two seats. It produced the fastest typical answer on the board and the cheapest results anywhere at $0.07 per correct answer. Kimi’s 4–1 head-to-head edge, however, keeps the open-weights title with Moonshot.

• Qwen 3.7 Max ticked up to 32%. It remains the budget model that somehow costs about twice as much per result as its budget rivals, at $0.34 per correct answer.

• GLM-5.2 improved for a third straight board: 2% → 15% → 22%. Watch this space. Eventually.

Why did the board move?

Every returning model scored higher than it did on the previous board. Today’s problem slate was easier, and our day-to-day comparability controls attribute that broad lift to the problems, not to a simultaneous improvement across every model.

The ordering is where the news lives.

Of the 36 order calls between returning models, 29 repeated (τ = 0.61). Most of the seven swaps came from two genuine movers:

1. GPT-5.5 climbed through the top four.

2. Opus 4.8 rebounded through the middle.

Sol also slipped below both.

Raw percentages are a model’s daily costume. Rankings and head-to-head results are the body underneath.

Why not trust the benchmarks you have already seen?

Public benchmarks eventually leak into training data. Scores inflate, and what once measured ability starts measuring memorization.

They are also snapshots: accurate on the day they are run, then progressively less representative as models, serving infrastructure, quantization, and inference behavior change.

We generate fresh, novel problems every day. No model can have trained on a problem that did not exist until that morning.

That lets us measure a model released on Thursday against six-week incumbents on identical footing. It also lets us retest the launch-day shine four days later.

Today’s board required approximately $155 in compute. But you can subscribe for free!

Subscribe now

What to do with today’s board

Using GPT-5.5? Stay put.

Today, it matched the most expensive model on the board at 60% of the realized cost and beat its own successor head-to-head. There is currently no demonstrated reason to migrate.

Paying Fable prices? Run Grok 4.5 alongside it for a week.

Grok cost $0.32 per correct answer, compared with Fable’s $1.43, while remaining statistically level. Let your own workload settle the argument.

Considering a move to GPT-5.6 Sol? Hold off.

Wait until it beats GPT-5.5 on this board for at least two consecutive days. Speed is currently its only proven advantage, and today you paid nearly twice as much per correct answer for it.

That is our read. Argue with it. Tomorrow’s board is the rebuttal.

TL;DR

Best closed model: Fable 5. It holds #1 for a seventh straight day, but only on a tiebreak. Grok 4.5 matched its 53% solve rate, and the two split their head-to-head 8–8.

Cheapest closed model: Grok 4.5. It delivered champion-level results at $0.36 per correct answer, roughly one-fifth of Fable’s $1.95.

Best open weights model: Kimi K2.7 Code. It reached 42%, statistically level with GPT-5.5, at just $0.15 per correct answer.

The board shook up. Only 23 of 28 pairwise orderings between returning models held from yesterday (τ = 0.64). Every swap came from three movers: Opus 4.8 falling, while Grok 4.5 and Kimi K2.7 Code climbed.

Today’s tidbit: Meta’s Muse Spark 1.1 tops the raw table at 75%, but it does not get the crown. Roughly 9 in 10 API requests failed on launch day, and Meta still billed for the failed calls. Count the full bill, and Muse cost about $25 per correct answer, the most expensive results on the board disguised behind its cheapest sticker.

Five things worth your attention today

1. Grok 4.5 catches the champion, and the price gap is a chasm

Yesterday, Grok debuted mid pack and we called it the value pick.

Today, it caught the leader outright.

Grok 4.5 and Fable 5 both solved 53% of the slate. On the problems where they disagreed, they split them 8–8 (p = 1.0).

You cannot get more tied than that.

Fable keeps the higher rank on Elo because it beat stronger opponents, but the money column is not close:

• Grok 4.5: $0.36 per correct answer

• Fable 5: $1.95 per correct answer

That is more than five times the results per dollar.

Grok also did something no budget priced model has managed on this board: it separated from the cheap pack.

It swept GLM 5.2 15–0 (p < 0.001), thumped Nemotron 3 Ultra 14–2 (p = 0.004), and beat Qwen 3.7 Max 15–5 (p = 0.04).

Fable’s consolation is speed. It was the fastest full slate model on the board, with a typical answer in about a minute and a half, compared with roughly four minutes for Grok.

If today repeats tomorrow, the daily driver question has a new default answer.

2. About Meta: Muse Spark’s debut mostly didn’t happen

Meta’s first paid model would’ve topped the table at 75%.

We are not crowning it.

Roughly nine out of every ten attempts came back as errors from Meta’s day one API and never produced an answer to score.

On the small fraction that got through, Muse was quick, with a typical answer in about a minute. But the confidence interval around its 75% score runs from approximately 30% to 95%.

That is statistics speak for: we know almost nothing yet.

You can see it in the Elo chart below, where Muse’s error bars dwarf everyone else’s.

There is also a launch day pricing trap.

The leaderboard’s cost column only meters answers that came back successfully. Meta, however, billed for the failed attempts too.

Count everything on the bill, and Muse’s correct answers cost about $25 each.

That makes them the most expensive results on today’s board by an order of magnitude, despite Muse wearing the cheapest sticker in the table.

Launch day API instability is normal, and these failures were on Meta’s side rather than the model’s. But until the API survives a full slate, Muse is neither the best model on this board nor remotely the cheapest.

3. GPT 5.6 Sol arrives: better than GPT 5.5, but pricier per result

OpenAI’s new flagship landed at #3 on day one.

It edged GPT 5.5 7–5 head to head (p = 0.77). That is directionally an upgrade, but statistically still a coin flip.

The catch is efficiency.

GPT 5.6 Sol carries the same sticker price as GPT 5.5, but it produced roughly 60% more output to get its answers. Its realized cost therefore came out higher:

• GPT 5.6 Sol: $2.10 per correct answer

• GPT 5.5: $1.44 per correct answer

Sol was quick about it, with a typical answer just over two minutes, compared with roughly five and a half minutes for GPT 5.5. It also completed the entire slate.

Against Fable 5, it is already statistically level, with a 7–9 split (p = 0.80).

If you are using GPT 5.5 today, Sol buys you a few more solves and a faster answer. It charges you for the extra words.

Watch this one. Debuts move.

4. Opus 4.8 fell five places, and every part of its line got worse

After spending two weeks at or near the top, Opus 4.8 dropped to 38% today, down from 48%.

It fell past GPT 5.5, Grok 4.5, and Kimi K2.7 Code on the way to seventh place.

Against Grok, it lost 5–11 on their discordant problems.

Opus was also the only model to leave problems unanswered today. That unfinished work still appeared on the bill.

Its $2.18 per correct answer in metered spend was the worst value among the models that completed most of their slate.

One brutal day is a data point, not a verdict. Fable had clear daylight over Opus yesterday, and the two remain statistically level today at 11–5 (p = 0.21).

But for the first time, the Anthropic premium bought seventh place results.

5. Kimi K2.7 Code just gate crashed the flagship club

Quietly, Kimi delivered the day’s biggest climb in raw score.

Kimi K2.7 Code jumped from 22% to 42%.

It finished statistically dead even with GPT 5.5, splitting their discordant problems 8–7 (p = 1.0). It also edged Opus 4.8 9–8 on the problems where the two models disagreed.

Its cost: $0.15 per correct answer, the second cheapest results anywhere on the board.

The rest of the bargain tier also had a day.

GLM 5.2 recovered from yesterday’s one solve catastrophe to reach 15%, while Nemotron 3 Ultra remained the absolute cheapest model per result at $0.12, with a typical answer in under two minutes.

But Kimi is the story.

If it holds anywhere near this line, “flagship results at 94% off” stops being a hypothetical.

Why the board moves at all

Eight models returned from yesterday, and today’s board genuinely reshuffled.

Of the 28 pairwise orderings between returning models, 23 repeated, producing a Kendall rank correlation of τ = 0.64.

All five swaps trace back to two movers:

• Opus 4.8 fell behind GPT 5.5, Grok 4.5, and Kimi K2.7 Code.

• Grok 4.5 climbed past GPT 5.5.

That is the useful kind of low stability day.

It was not noise smearing the entire board. It was two models genuinely trading altitude.

Providers update serving stacks silently. Models drift. Inference is nondeterministic. Our day to day comparability controls put today’s swings on the models’ side of the ledger.

That is exactly why we re-measure every morning instead of trusting last week’s board.

Why not just trust the benchmarks you’ve already seen?

Public benchmarks leak into future training runs.

Scores inflate, and eventually the benchmark starts measuring memorization rather than ability.

They are also snapshots: accurate on the day they are taken, and less useful every day afterward.

We generate fresh, novel problems every day. No model can have trained on a problem that did not exist until that morning.

That is also why debutants such as GPT 5.6 Sol, tested only hours after general availability, can be evaluated on exactly the same footing as models we have tracked for weeks.

Today’s board required about $242 in compute.

Start your day with the right model

The best model is a daily fact, not a permanent one.

Today, a six day champion got caught by a challenger charging a fifth as much. A new flagship debuted mid pack. Last week’s #2 fell five places.

Tomorrow decides which of those moves was real.

Subscribe, and each morning you’ll get the day’s best model, plus the fastest and cheapest way to get that performance, in your inbox before you start work.

Subscribe now

Grok 4.5: ~70% cheaper than Opus per correct answer (69.6% exactly), but only ~53% cheaper than GPT-5.5 per correct answer, because GPT-5.5 is itself much cheaper per result than Opus today.

The field compresses, the order holds, and the value math resets.

Today’s fresh problems ran hard.

Every returning model scored lower than yesterday, yet the order at the top did not move an inch. That tells you the drop came from the day’s exam, not the students taking it.

Fable 5 held #1 for a sixth straight day at 57%.

The debutant, SpaceXAI’s Grok 4.5, landed mid-pack at 35%. It was statistically level with every flagship except Fable, while costing 79% less than the priciest model on the board.

And GLM-5.2 solved exactly one problem all day.

Five things are worth your attention today.

1. Six straight days of Fable 5, on the day the test pushed back hardest

On a slate that knocked double digits off every flagship’s score, Fable 5 kept roughly the same daylight it had yesterday.

It finished #1 at 57%, nine points clear of Opus 4.8.

But the statistics remained humble.

Fable’s edge over Opus 4.8 was 9–5 on the problems that separated them, with p=0.42, effectively a coin flip. Against GPT-5.5, it went 10–5, with p=0.30.

In fact, the only matchup anywhere in the top six that cleared statistical significance was Fable over the newcomer Grok 4.5: 13–4, with p=0.049.

Fable also finished everything it started. It left zero problems unanswered.

At $1.61 per correct answer, Fable now sits between Opus 4.8 at $1.88 and GPT-5.5 at $1.21 per result, despite carrying the highest sticker price.

The crown holds. The moat is still measured in head-to-head wins, not certainty.

2. Grok 4.5’s debut resets the value math

On its first day on the board, Grok 4.5 immediately became the price-performance story.

At a 35% solve rate, it was statistically indistinguishable from:

• Opus 4.8

• GPT-5.5

• GPT-5.4

• Opus 4.6

That is every flagship except Fable 5.

Grok delivered those results at $0.57 per correct answer. That is less than half of GPT-5.5’s $1.21 and roughly one-third of Fable’s $1.61.

It also answered every problem it was given. There were no timeouts and no abandoned attempts. A typical answer took about five minutes.

For workloads that live in the “very good, but not necessarily the absolute best” band, today’s board says Grok 4.5 is where the dollars go furthest among the major models.

3. Opus 4.6 is on the board too, and it is a cautionary tale

The other newcomer had a much rougher landing.

Opus 4.6 matched GPT-5.4’s solve count, but one-quarter of its attempts ended without an answer at all.

That was by far the worst reliability on the board, on a day when seven of the ten models finished every problem they attempted.

Between its near-flagship sticker price, only 8% below the most expensive model, and the unanswered problems, Opus 4.6 reached $2.27 per correct answer.

That was the most expensive result anywhere on today’s board.

It was also cost-dominated: Grok 4.5 matched it on results while beating it decisively on price.

For Anthropic users, Opus 4.8 is the better buy. The generation old discount is not much of a discount.

Sonnet 5, yesterday’s reliability story, sat today’s board out.

4. GPT 5.4’s discount is now slow and expensive

GPT-5.4 held its place alongside the flagships at 38%, but it was also the slowest model on the board.

A typical answer took around twenty minutes, compared with roughly three minutes for GPT-5.5.

The verbosity that keeps erasing GPT-5.4’s sticker discount also got worse. It produced roughly twice as much output as GPT-5.5.

As a result, the supposedly cheaper model landed at $1.48 per correct answer, compared with $1.21 for GPT-5.5.

The conclusion is the same as yesterday, now with a stopwatch attached:

Between the two GPT models, buy 5.5.

5. GLM 5.2 did not slump. It vanished.

Yesterday, GLM-5.2 bounced from 8% to 20%.

Today, it solved one problem.

It was shut out by models in its own price neighborhood:

• 0–13 against Grok 4.5

• 0–8 against Kimi K2.7 Code

• 0–7 against Qwen 3.7 Max

• 0–7 against Nemotron 3 Ultra

Fable 5 swept it 22–0.

When the problems become difficult, GLM 5.2 currently has nothing to offer. A 97% sticker-price discount is not a bargain when the model solves 2% of the test.

The budget money belongs where it went yesterday.

Kimi K2.7 Code held up best in the bargain bin, with a 22% solve rate and a cost of $0.24 per correct answer.

Nemotron 3 Ultra remained the cheapest source of correct answers anywhere on the board at $0.15 per result. It answered every problem and typically finished within four minutes.

Why the board moves at all

Nine models returned from yesterday.

Every one of them scored lower today, yet the top four finished in exactly the same order.

That pattern is the tell.

Nothing suddenly decayed overnight. Fresh problems differ from day to day, and today’s set was simply harder. Our day-to-day comparability controls also place the change on the problems’ side of the ledger, not the models’.

That is why the numbers to trust across days are the rankings and head-to-head results, not the raw percentages alone.

A board that does not reshuffle on a brutal day is evidence that the ranking is real.

Why not just trust the benchmarks you have already seen?

Public benchmarks eventually leak into future training runs.

Scores inflate, and the benchmark gradually begins measuring memorization rather than ability.

They are also snapshots. A result may be accurate on the day it is measured, then become less useful with every model and infrastructure update that follows.

We generate fresh, novel problems every day.

No model can have trained on a problem that did not exist until that morning.

That also means a newcomer such as Grok 4.5 is evaluated on the same footing as models we have tracked for weeks. Its first day on the board uses the same fresh test as everyone else.

Today’s board required about $168 in compute.

Subscribe now

GPT 5.6 is expected later today and will be included in the next post.

Start your day with the right model

The best model is a daily fact, not a permanent one.

Today, the champion made it six days in a row, a debutant rewrote the value math, and last week’s comeback story forgot how to answer.

Tomorrow will tell us which of those signals was real.

The crown holds for a fifth day, but the race at the top is suddenly a coin flip

Yesterday, Fable 5 swept Opus 4.8 7–0 on the problems that separated them.

Today, that same matchup went 5–3, effectively a coin flip (p=0.73).

Fable 5 holds the top spot for a fifth straight day at 78%, but Opus 4.8 climbed to 72%, up from 65%, while GPT-5.5 surged to 72%, up from 60%. GPT-5.4 also climbed all the way into the flagship pack.

The top four models are now statistically inseparable.

The model that is separated, in the wrong direction, is Sonnet 5.

Four things worth your attention today

1. The crown held. The moat didn’t.

Fable 5’s fifth day at #1 was its most contested yet.

Its edge over Opus 4.8 went from a clean sweep to 5–3, and its matchup with GPT-5.5 is just as much of a tie.

What still separates Fable from the pack it can no longer statistically beat?

It was the fastest flagship by a wide margin. Its typical answer arrived in under two minutes, while the other flagships took between three and seven minutes.

At $1.20 per correct answer, it was also cheaper per result than Opus 4.8 at $1.23, despite carrying the highest sticker price on the board.

If today’s board is your buying guide, Fable is still the pick at the top. Just not by the margin yesterday suggested.

2. GPT-5.4 crashed the flagship party

The quiet story of the day is GPT-5.4.

It jumped to 68% and is now statistically indistinguishable from Opus 4.8 and GPT-5.5, while carrying a sticker price 44% below Fable 5.

There is one caveat before you swap it in everywhere: its famous discount keeps getting eaten by verbosity.

GPT-5.4 produced roughly twice as much output as GPT-5.5 today. Its realized cost landed at $0.77 per correct answer, compared with $0.73 for GPT-5.5.

The half-price model was, per successful result, slightly more expensive.

Between those two, buy GPT-5.5. Against Opus 4.8 at $1.23 per solve, either GPT is the value call today.

3. Sonnet 5 is the odd flagship out, and it’s a reliability story

Sonnet 5 landed at 50%, and today the gap above it is no longer statistical noise.

It lost its head-to-head matchups against:

• Fable 5: 12–1 (p=0.003)

• Opus 4.8: 11–2 (p=0.022)

• GPT-5.5: 1–10 (p=0.012)

The failure mode matters as much as the final score.

More than a quarter of Sonnet 5’s attempts ended with no answer at all, by far the worst rate on the board. Every budget model completed everything it started.

A relatively affordable premium model that silently fails to answer is a worse deal than its $0.61 per solve suggests.

If you are paying mid-tier money, GPT-5.4 occupies roughly the same cost-per-result neighborhood while delivering results a full tier above.

4. The bargain bin reshuffled again

GLM-5.2 bounced off yesterday’s floor, climbing from 8% to 20%.

It still finished dead last, got shut out by its own price peers, and was the slowest model on the board:

• 0–8 against Qwen 3.7 Max

• 0–9 against Kimi K2.7 Code

The steadier story is Kimi K2.7 Code.

Across three boards, it has posted three consecutive improvements:

32% → 40% → 42%

It also recorded zero unanswered problems and delivered correct answers for $0.12 each.

Right now, Kimi is the most dependable sub-$5 model we track.

Nemotron 3 Ultra cooled from 44% to 38%, but retained both crowns that matter for high-volume work:

• Fastest model on the board, with a typical answer in about 70 seconds

• Cheapest correct answers anywhere, at $0.08 per solve

Why does the board move at all?

No model on this board shipped a version bump in the past 24 hours.

Yet a 7–0 sweep became a coin flip, GPT-5.5 gained 12 percentage points, and GLM-5.2 recovered half of Monday’s collapse.

Providers silently change serving infrastructure. Capacity and routing shift. Quantization, inference settings, and nondeterminism do the rest.

Our day-to-day comparability controls put these swings on the models’ side of the ledger, not ours.

Why not just trust the benchmarks you’ve already seen?

Public benchmarks leak into future training runs. Scores inflate and gradually begin measuring memorization rather than ability.

They are also snapshots: accurate on the day they are measured, and less representative with every day that passes.

We generate fresh, novel problems every morning. No model can have trained on a problem that did not exist until today.

A strong score here reflects genuine ability, not recall.

We then run the entire field and evaluate capability alongside cost and speed.

Today’s board required roughly $137 in compute.

Subscribe now

Start your day with the right model

The best model is a daily fact, not a permanent one.

Today, the champion kept the crown but lost the moat. A discounted flagship earned its seat at the top table. A premium model forgot to answer.

Tomorrow will tell us which of those signals was real.