Field intelligence for AI-first professionalsVol. II · Nº 195 · Monday, November 16, 2026
← Phantom Notes
Nº 187AI ModelsNovember 13, 20263 min read

Gemini 4 Carbon Tops MathArena at 85.4%. Its Lead Is Inside the Error Bars, and Each Problem Costs 15 Times What GPT-6.1 Sol Charges.

MathArena, the ETH Zurich and INSAIT competition-math board, added Gemini 4 Carbon and Claude Fable 5.2 on November 13. Carbon at high leads at 85.4% plus or minus 3.0 with an expected cost of $15.70 per problem; Fable 5.2 at max scores 82.8% plus or minus 5.3 at $40.20; GPT-6.1 Sol at max scores 80.6% plus or minus 2.7 at $1.07. The three intervals overlap, the effort settings differ, and Carbon's most common failure was running out the clock, once with a script that ran more than eleven hours. The table renders by JavaScript, so the figures come from The Decoder's copy, matched by independent readers on X. Carbon's API price doubles on December 1, and which rate MathArena used is not stated in what we read. Google has not announced the ranking.

By T.W. Ghost


MathArena, the competition-math leaderboard run by ETH Zurich's SRI Lab and INSAIT, posted an update today, November 13, 2026: "We added Claude Fable 5.2 and Gemini 4 Carbon results." On the overall table, Gemini 4 Carbon at high leads at 85.4%, plus or minus 3.0, with an expected cost of $15.70 per problem. Claude Fable 5.2 at max is second at 82.8%, plus or minus 5.3, at $40.20. GPT-6.1 Sol at max is third at 80.6%, plus or minus 2.7, at $1.07. The best open model is Kimi K3.1 Pro at max, at 54.3%, plus or minus 5.8, for $3.70.

Three catches sit on top of the ranking. The three intervals overlap, so the order of the top three is not settled by this table. The settings differ: Carbon ran at high, the other two at max. And the cost column is a spread of 15 times between Carbon and Sol and about 38 times between Fable and Sol, for a few points of accuracy inside the error bars. One more: MathArena's overall table renders by JavaScript, and we could not read it directly. The scores and costs here are The Decoder's copy of the table, which several independent readers on X matched. We read the MathArena update line, The Decoder's piece, Google's pricing page, and the X checks. Here is what held up.


The Table, as Relayed

Model (effort)ScoreExpected cost per problemWho
Gemini 4 Carbon (high)85.4% ± 3.0$15.70 ± $1.33MathArena table via The Decoder; matched by X readers
Claude Fable 5.2 (max)82.8% ± 5.3$40.20 ± $2.59same
GPT-6.1 Sol (max)80.6% ± 2.7$1.07 ± $0.018same
Kimi K3.1 Pro (max), best open54.3% ± 5.8$3.70same

The lead is inside the error bars. 85.4 plus or minus 3.0, against 82.8 plus or minus 5.3, against 80.6 plus or minus 2.7. A reader who wants a winner gets Carbon; a reader who wants a ranking gets three models that overlap.

High against max. Carbon's number is at its high setting; Fable's and Sol's are at max. Whether Carbon at a higher effort would score more, or Fable at high would cost less, is not on the table.

Fifteen times the price for the top row. $15.70 against $1.07 per problem, and Fable's $40.20 is about 38 times Sol. The Decoder relays that Carbon's most common failure was the time limit, and that one script ran unchecked for more than eleven hours, which is where the cost goes.

The price doubles December 1. Carbon's API rate is $2 and $10 per million through November 30 and $4 and $20 from December 1. MathArena does not say, in what we read, which rate its expected cost uses; if it is the intro rate, the gap to Sol widens next month.


Verified vs Unconfirmed: The Scorecard

ClaimVerdict
MathArena added Fable 5.2 and Carbon on Nov 13, 2026Verified (matharena.ai update line)
Scores, intervals, and costs for Carbon, Fable 5.2, Sol, Kimi K3.1 ProRelayed (The Decoder's copy of the table; matched by independent X readers)
Carbon's time-limit failures; eleven-hour scriptPress (The Decoder relaying MathArena)
Four runs per problem; aggregate of non-deprecated competitionsUnverified (X reader of the site; not confirmed in our read)
ArXivMath October figures: Carbon 93.62% at $5.29; Fable 93.62% at $22.28; Sol 90.43% at $1.21Unverified (X only)
Carbon $2 / $10 through Nov 30; $4 / $20 from Dec 1Verified (Gemini API pricing)
Which rate MathArena's cost usesNot stated in what we read
Google has announced the rankingNo (press and X)

Who Should Use It

Use Carbon for competition-style math if the last few points matter and $15 a problem does not, and run it at a setting you have measured, because the clock is where it fails.

Skip the ranking as a verdict; three overlapping intervals and two different effort settings make it a tie with a price column, and the price column favors Sol by a wide margin.

The frontier is elsewhere: on Artificial Analysis's general Index the order is Fable 5.2 at 65 ahead of Carbon at 60, which is a reminder that one board is one board. Our model comparison has the current standings side by side.


Sources


*Not sure whether a math board tells you anything about your workload? Take the free 2-minute quiz and get matched. Then read about the model that pays $1.07 for third place: GPT-6.1 Sol.*

Which model should you be using?

Three minutes, twelve questions, one defensible answer.

Take the quiz →