SuperIntelligence Infrastructure 7.2%reading
modelsGoogleLayer lab

Google's Gemini 4 Carbon Tops MathArena, Billing 15 Times GPT-6.1 Sol at the Same Token Price

MathArena ranks Google's Gemini 4 Carbon first at 85.4 percent, yet Jasper Dekoninck's own tables show a tie with Claude Fable 5.2 on research problems and a bill 15 times GPT-6.1 Sol's.

▲ +0.3 Capability reported Reading after Nov. 16, 2026: 7.8

By Ryan Elliott Dennis · 13 sources · 8 min read

Glass office building with the Google logo behind trees at the Mountain View headquarters
Google's headquarters in Mountain View, California, photographed in August 2017; Google opened Gemini 4 Carbon to paying developers on Nov. 10, 2026. Photo: David Nagle, Wikimedia Commons, CC BY-SA 4.0

Google's Gemini 4 Carbon took first place on MathArena, the math evaluation built by the SRI Lab at ETH Zurich and INSAIT, when the board added its results on Nov. 13, 2026 2. Carbon scored 85.4 percent, plus or minus 3.0 points, at an expected cost of $15.70 a problem. OpenAI's GPT-6.1 Sol, the model it displaced, scored 80.6 percent at $1.07 1. MathArena billed both at $2 per million input tokens and $10 per million output tokens 310.

Why should a reader outside a math department care? A math answer is right or wrong, so an outside grader can check a model's work exactly, and this ledger counts a capability gain when an outsider measures it. MathArena is that outsider. Its table also prices the gain: 4.8 points for 15 times the inference cost, the money spent running a model on each task.

Gemini 4 Carbon's MathArena lead adds 0.3 to capability

The move
Capability steps up 0.3 at reported confidence, the floor of the 0.3 to 0.8 band: MathArena measured Carbon 4.8 points ahead of GPT-6.1 Sol, while Fable 5.2's interval reaches above Carbon's score and the bill runs 15 times Sol's.
The data5 rows · sources
MeasureValue
Reading before this day7.5
This piece's move+0.3 (Capability, reported)
Band for reported evidence0.3 to 0.8
Reading after the day7.8
Distance to 10092.2

What did MathArena measure?

MathArena runs each model four times on every problem and reports the average score beside the average cost of a single run 2. "We use new problems to measure reasoning on questions that models have not seen during training," the site says. Expected performance, its headline figure, averages scores across every current competition and fills gaps with a statistical estimate wherever a model skipped one.

Three reasoning models lead the table, each listed with the reasoning-effort setting MathArena ran it at. Carbon, at high effort, holds 85.4 percent. Anthropic's Claude Fable 5.2, at max effort, holds 82.8 percent, plus or minus 5.3 points. GPT-6.1 Sol follows at 80.6, plus or minus 2.7. Moonshot AI's Kimi K3.1 Pro, the best open model, trails at 54.3 percent.

Is the order of the top two settled? Read the intervals. Carbon's range runs from 82.4 to 88.4 and Fable's from 77.5 to 88.1, by desk arithmetic, so Fable's upper bound sits above Carbon's score. Tbreak made the same point on Nov. 14, writing that Carbon's range overlaps both rivals' 12. Overlap alone leaves the question open, because only a test on paired, problem-by-problem results would settle it, and the summary table stops short of one.

Monthly sheets sharpen the picture. On ArXivMath, the set drawn from recent research papers, Carbon and Fable posted identical scores in both months they share: 96.23 percent in September and 93.62 percent in October 311. Sol edged Carbon on September's BrokenArXiv, 82.70 to 82.39 10. BrokenArXiv hands a model a subtly false statement and scores whether the model says so 5. Carbon's widest margin came on ArXivLean, which asks for formal proofs in the Lean language: 62.26 percent against Sol's 35.85.

Stone steps leading to the decorated facade and wooden door of ETH Zurich's main building
The main building of ETH Zurich, photographed in September 2022; the university's SRI Lab builds MathArena with INSAIT. Photo: Leonhard Lenz, Wikimedia Commons, CC0

MathArena expected performance, Nov. 13: Carbon first by 2.6 points

Compared
Gemini 4 Carbon leads at 85.4 percent, 2.6 points over Claude Fable 5.2 and 4.8 over GPT-6.1 Sol; Fable's interval of plus or minus 5.3 points reaches above Carbon's score, and Kimi K3.1 Pro leads open models at 54.3.Sources [2]
The data4 rows · sources
Unit: percent
ItemValueSource
Gemini 4 Carbon
high, ±3.0
85.4[2]
Claude Fable 5.2
max, ±5.3
82.8[2]
GPT-6.1 Sol
max, ±2.7
80.6[2]
Kimi K3.1 Pro
max, best open model
54.3[2]

Where does Google say Carbon is strongest?

Koray Kavukcuoglu, senior vice president at Google DeepMind and Google's chief AI architect, announced general availability on Nov. 10 4. "Across independent benchmarks designed to test model performance in real workflows, Carbon pairs frontier accuracy with practical production economics," he wrote. His post cites first place on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, each at under half the cost per task of the next best outside model. MathArena's result arrived three days later, and it supports the first half of his sentence while the second half meets a $15.70 bill.

Kavukcuoglu's post also helps explain the Lean result. Google researchers built an agent framework called AlphaProof Code, he wrote, and its agents used Carbon to generate millions of lines of Lean while proving security properties of the Signal messaging protocol. Read as strategy, Google aimed Carbon at formal proof, and the ArXivLean sheet is where its margin is widest: 26.4 points over Sol 310.

He named the cost of that design himself. Carbon, he wrote, is "using more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops." Gemini 4 Argon is the model Google still recommends for most software engineering and knowledge work.

Entrance of the 6 Pancras Square office building in London that houses Google DeepMind
The entrance of 6 Pancras Square in London, home to Google and its DeepMind unit, photographed in December 2017. Photo: Gciriani, Wikimedia Commons, CC BY-SA 4.0

Why does one token price produce a bill 15 times larger?

Start with the price list. Simon Willison, the developer behind Simon Willison's Weblog, tested Carbon on launch day 6. "This is Google's best ever Gemini, priced at $2/million input and $10/million output at launch (that's GPT-6.1 Sol / Claude Sonnet 5.5 prices) but scheduled to double in price to $4/$20 on December 1st (same as Opus 5.5)," he wrote. The Decoder reported the same $2 and $10 for Sol at its release on Sept. 29 13.

So the gap is volume. On BrokenArXiv, Carbon wrote 539,577 output tokens per problem on average and Sol wrote 80,413, close to seven times as many 310. ArXivLean widened it: 1,064,444 tokens and $45.11 for Carbon, 59,111 tokens and $1.16 for Sol.

Theo Browne, a developer who reviewed Carbon on his t3.gg podcast, stated the principle, according to BigGo Finance's account of the episode 7. "The token cost does not matter. The cost per task is what matters, especially when you consider how many tokens are being used in a given task," Browne said. MathArena's cost column measures exactly that.

One date changes the arithmetic. Google's introductory price ends on Nov. 30, and the standard price of $4 and $20 applies afterward 4. If Carbon's token counts hold, its MathArena cost doubles to about $31.40 a problem, an estimate by desk arithmetic, or roughly 29 times Sol's. Fable 5.2 already costs $40.20, which is 2.6 times Carbon's bill for a score 2.6 points lower 2.

MathArena expected cost per problem: $15.70 for Carbon, $1.07 for Sol

Compared
MathArena bills Carbon and GPT-6.1 Sol at one token price, $2 and $10 per million, yet Carbon's expected cost per problem is $15.70 against Sol's $1.07, and Fable 5.2 costs $40.20 for a lower score.Sources [2]
The data4 rows · sources
Unit: US$ per problem
ItemValueSource
Claude Fable 5.2
82.8 percent
40.2[2]
Gemini 4 Carbon
85.4 percent
15.7[2]
Kimi K3.1 Pro
54.3 percent
3.7[2]
GPT-6.1 Sol
80.6 percent
1.07[2]

Why is the step only 0.3?

Capability moves up 0.3 on the ledger at reported confidence, the floor of the 0.3 to 0.8 band. An outside board measured a public model ahead of the previous leader, which earns the step. Four facts hold it at the floor: contamination, the clock, the intervals and a second board.

First, the test's authors concede a weakness. Jasper Dekoninck of ETH Zurich, first author of the MathArena paper, wrote with four co-authors on Sept. 15 that the team had stopped excluding arXiv papers that mention AI use 59. "This creates a potential concern: we are building benchmarks from questions that may have been answered using the models themselves, so the model used most in practice may have an advantage," they wrote. Their next lines go further: "Unfortunately, there is no good way around this issue." The ledger's method reserves confirmed confidence for a measured result on an uncontaminated test, and MathArena's builders decline to promise one.

Second, Carbon runs long. "The model also almost never gives up on a task, a trait that can be read as either positive or negative," Matthias Bastian wrote in The Decoder on Nov. 14 1. Its most common failure is exceeding the time limit, he reported, and one run launched a script that ran for more than eleven hours while Carbon left its progress unchecked. MathArena's September post set the limit at 12 hours and $100 per attempt, so an eleven-hour script uses more than 90 percent of the allowance on one unwatched step 5. The ledger's definition asks for weeks of unsupervised work at expert reliability, and a model that loses track of a 12-hour clock sits far from that standard.

Third come the intervals already described, which leave Fable 5.2 within reach of first place. Fourth, a second board ranks the pair in reverse. Artificial Analysis puts Fable 5.2's High tier at 62 on its index and Carbon at 60, The Decoder reported on Nov. 10, with Carbon at $7.76 per solved task and Fable's High tier at about $9 8.

Against all four stands the measured gap over Sol. Carbon's lower bound of 82.4 sits 0.9 points under Sol's upper bound of 83.3, a far thinner overlap than the one with Fable, and the Lean margin is 26.4 points. That evidence earns a step on the capability component, and the four limits keep the step small.

What MathArena measured for Carbon, and what stays open

The record
Solid tiles come from MathArena's published tables; dashed tiles show why capability steps up only 0.3: contamination its authors concede, a lead over Fable 5.2 inside the intervals, timeouts and a price that doubles Dec. 1.Sources [1] [2] [3] [4] [5] [8] [10] [11]
The data9 rows · sources
ColumnItemSource
MeasuredCarbon first at 85.4 percent, 4.8 points over GPT-6.1 Sol[2]
MeasuredArXivLean: Carbon 62.26 percent against Sol's 35.85[3] [10]
MeasuredEach model runs four times on every problem[2]
MeasuredCarbon and Sol billed at $2 and $10 per million tokens[3] [10]
OpenAuthors concede questions may already sit in training data[5]
OpenCarbon and Fable 5.2 tie on ArXivMath in September and October[3] [11]
OpenCarbon's most common failure is exceeding the time limit[1]
OpenGoogle's price doubles to $4 and $20 after Nov. 30[4]
OpenArtificial Analysis ranks Fable 5.2 High at 62, Carbon at 60[8]

What would each side need to be true?

Kavukcuoglu's claim of practical economics holds if Carbon keeps first place after Dec. 1 at a price research groups accept, and if a paired test separates it from Fable. Dekoninck's caution holds if a month of fresh problems reorders the top three. The MathArena paper itself warns that static benchmarks are "quickly saturated, and rarely updated," which is why Dekoninck's team keeps designing new benchmarks as models improve 9.

For a mathematician or a research manager choosing a model for proof work, MathArena's table gives a rule: read the cost column beside the score column. A reader tracking the buildout should count Carbon's lead over Sol as measured, its lead over Fable as unsettled, and $15.70 as a November price.

Kavukcuoglu claims practical economics; Dekoninck concedes a benchmark weakness

Both sides
Kavukcuoglu of Google DeepMind says Carbon pairs frontier accuracy with practical economics, while Dekoninck of MathArena concedes the most-used model may hold an advantage; the evidence leans slightly toward a real, small capability gain.Sources [4] [5]
The data2 rows · sources
SideWhoClaimSource
ForKoray KavukcuogluAcross independent benchmarks designed to test model performance in real workflows, Carbon pairs frontier accuracy with practical production economics.[4]
AgainstJasper DekoninckWe are building benchmarks from questions that may have been answered using the models themselves, so the model used most in practice may have an advantage.[5]

By the numbers

  • 85.4 percent, plus or minus 3.0: Gemini 4 Carbon's expected performance on MathArena, first of all models listed 2.
  • $15.70 against $1.07: expected cost per problem for Carbon and for GPT-6.1 Sol, a ratio of roughly 15 to 1 1.
  • Claude Fable 5.2 scored 82.8 percent, plus or minus 5.3, at $40.20 a problem 2.
  • 96.23 percent: the ArXivMath score Carbon and Fable 5.2 each posted for September 2026 311.
  • Carbon wrote 1,064,444 output tokens per ArXivLean problem and scored 62.26 percent; Sol wrote 59,111 and scored 35.85 310.
  • Twelve hours and $100: the time and cost limits MathArena set per attempt in September 5.
  • Nov. 30: the last day of Google's introductory price of $2 and $10 per million tokens, before $4 and $20 apply 4.
  • Kimi K3.1 Pro, the best open model, scored 54.3 percent at $3.70 a problem 2.

What to watch

MathArena publishes new ArXivMath and BrokenArXiv sheets each month, and the November sheets will show whether Carbon's lead over Fable 5.2 survives fresh problems 3. Google's standard price takes effect on Dec. 1, so the board's cost column for Carbon should rise if its token counts hold 4. A paired comparison from Dekoninck's team, or a published timeout rate for each model, would confirm this move. Evidence that leaked problems lifted scores, or a reordered top three, would reverse it.

Sources

  1. 1Gemini 4 Carbon takes the MathArena crown but costs 15 times more than GPT-6.1 Sol, The Decoder, Matthias Bastian, Nov. 14, 2026
  2. 2MathArena: Model performance across MathArena, MathArena, SRI Lab at ETH Zurich and INSAIT, Nov. 13, 2026
  3. 3Gemini 4 Carbon (high), model page, MathArena, SRI Lab at ETH Zurich and INSAIT, Nov. 13, 2026
  4. 4Gemini 4 Carbon is now available, Google, Koray Kavukcuoglu, Nov. 10, 2026
  5. 5ArXivMath and BrokenArXiv: Harder Problems and Revised Grading, MathArena, Jasper Dekoninck, Tim Gehrunger, Ivo Petrov, Kári Rögnvaldsson and Martin Vechev, Sept. 15, 2026
  6. 6llm-gemini 0.36, Simon Willison's Weblog, Simon Willison, Nov. 10, 2026
  7. 7Gemini 4 Carbon Halves Google's Hallucinations, Review Finds, but Still Won't Beat Anthropic's Opus, BigGo Finance, Nov. 11, 2026
  8. 8Gemini 4 Carbon arrives just in time to feel like yesterday's news in the fastest model cycle yet, The Decoder, Matthias Bastian, Nov. 10, 2026
  9. 9Beyond Benchmarks: MathArena as an Evaluation Platform for Mathematics with LLMs, arXiv, Jasper Dekoninck, Nikola Jovanović, Tim Gehrunger, Kári Rögnvaldsson, Ivo Petrov, Chenhao Sun and Martin Vechev, May 15, 2026
  10. 10GPT-6.1 Sol (max), model page, MathArena, SRI Lab at ETH Zurich and INSAIT, Nov. 13, 2026
  11. 11Claude-Fable-5.2 (max), model page, MathArena, SRI Lab at ETH Zurich and INSAIT, Nov. 13, 2026
  12. 12Gemini 4 Carbon tops MathArena, at 15 times GPT-6.1 Sol's run cost, Tbreak, Tbreak Editor, Nov. 14, 2026
  13. 13GPT-6.1 Sol comes close to Astra at a fifth of the price, The Decoder, Maximilian Schreiner, Sept. 29, 2026