Close-up of hands using a calculator and writing equations on a notebook, perfect for math concepts
Models & capabilityIndependently verified

MathArena puts Gemini 4 Carbon first at 85.4 percent, at about 15 times the cost of GPT-6.1 Sol

By Yash Malviya·

Photo: https://kaboompics.com/ / Pexels

MathArena's expected-performance table, read on November 17, 2026, ranks Google's Gemini 4 Carbon first on math, but the gap to OpenAI's GPT-6.1 Sol is 4.8 points for roughly $14.60 more per problem.

MathArena, the benchmark project that scores AI models on fresh math competitions, now has Google's Gemini 4 Carbon in first place. On the table we read on November 17, 2026, Carbon (high) has an expected performance of 85.4 percent, plus or minus 3.0. Anthropic's Claude Fable 5.2 is second and OpenAI's GPT-6.1 Sol third.

The ranking is real. The interesting part is the price tag and the error bars.

The table

MathArena's own page lists, for non-deprecated competitions, these top entries:

  • Gemini 4 Carbon (high), Google, released November 10, 2026: 85.4 percent (plus or minus 3.0), $15.70 (plus or minus $1.33) expected cost per problem.
  • Claude Fable 5.2 (max), Anthropic, November 9, 2026: 82.8 percent (plus or minus 5.3), $40.20 (plus or minus $2.59).
  • GPT-6.1 Sol (max), OpenAI, September 29, 2026: 80.6 percent (plus or minus 2.7), $1.07 (plus or minus $0.018).
  • Best open model, Kimi K3.1 Pro (max), Moonshot AI, October 23, 2026: 54.3 percent (plus or minus 5.8), $3.70.

The page says the expected performance figures are estimates. They combine observed results with item-response modeling for questions a model has not answered, and each carries a confidence interval.

What the intervals say

The headline gap between Carbon and Sol is 4.8 points. The intervals are plus or minus 3.0 and plus or minus 2.7. Those ranges nearly touch, so Carbon's lead over Sol is probable rather than certain. The gap to Fable 5.2 is 2.6 points against an interval of 5.3 on Fable's side, which means second and first are close to a statistical tie.

The more robust finding is the jump to a new tier below them: the best open model is more than 30 points behind. On math, closed frontier models are far ahead of open ones, at least on this benchmark.

Teenager in hoodie writing math formulas on a blackboard indoors, showcasing problem-solving skills
A blackboard of equations, the kind of competition problems MathArena uses. Photo: https://kaboompics.com/ / Pexels

The cost gap

The Decoder, reporting on November 14, 2026, says Carbon costs roughly 15 times as much as GPT-6.1 Sol under the reasoning settings used in testing, and Fable 5.2 about 40 times as much. Our arithmetic from MathArena's table agrees: $15.70 divided by $1.07 is about 14.7, and $40.20 divided by $1.07 is about 37.6.

Put another way, buying the 4.8 extra points from Carbon over Sol costs about $14.63 more per problem. On a batch of a thousand problems, the table implies roughly $15,700 for Carbon against $1,070 for Sol, though real costs depend on settings, caching and discounts.

Carbon's behavior explains part of the bill. The Decoder reports that its most common failure is exceeding the time limit, that it almost never gives up on a task, and that in one case it ran a script for more than eleven hours without checking on progress. A model that does not quit is useful when the answer matters and expensive when it loops.

What this does and does not show

MathArena is one of the more careful public benchmarks, because it uses recent competitions that models could not have seen in training. Still, it measures competition math. It says little about coding agents or long tasks, where our own coverage of Carbon's launch benchmarks lists other claims, some of them Google's own.

Fable 5.2's price on MathArena also differs from its list price, because the table's figure is cost per problem at the highest reasoning setting, not per token. Our Fable 5.2 launch piece covers the token prices.

How MathArena works

MathArena's page says its headline score is the expected performance across non-deprecated competitions, and lists a wide set: Kangaroo 2025, AIME 2025 and 2026, three HMMT contests, BRUMO, SMT, CMIMC, Apex, USAMO 2025 and 2026, IMO 2025, IMC 2025, Miklos Schweitzer 2025 and Putnam 2025, plus Project Euler and newer research-style sets such as LeanLean, ArXivLean and BrokenArXiv. Some sets are marked deprecated and excluded. Because the data comes from recent contests, the risk of the model having memorized the answers is lower than on older benchmarks, though not zero.

The independence matters here. MathArena is not run by any of the three labs on the table, which makes it a more useful referee than a launch post. Google's own Gemini 4 Carbon post lists a different kind of claim, such as 92 percent on its Agentic Honesty Benchmark versus 70 percent for Argon, and a first-place rank on CursorBench 4.0, all company-reported. The math table does not test any of those.

One more caution. The table reads the highest reasoning setting for each model: Carbon "high," Fable 5.2 and GPT-6.1 Sol "max." A different setting would change both the score and the cost column, and a team that runs models at lower settings in production should not assume the same ordering holds.

The practical reading for a buyer is about cost per correct answer. Using the table's own figures, Carbon's $15.70 at 85.4 percent and Sol's $1.07 at 80.6 percent differ by an order of magnitude in price for a gap that sits inside the error bars. A team that needs the last few points on hard proofs may justify Carbon. A team doing routine math at volume almost certainly will not.

Our take

A fair reading: Carbon is probably the best math model right now, Sol is the value pick by an order of magnitude, and Fable 5.2 is within the noise of both. If you run math or proof workloads at scale, test Sol first and escalate hard problems to Carbon. If you are choosing a lab on one leaderboard, you are choosing on a four-point lead that could vanish on the next update.

What to watch: whether OpenAI's newer models, which the Decoder says are not yet publicly available, appear in the table, and whether Google's introductory Carbon pricing, which our earlier reporting says rises after November 30, changes the cost column.

Frequently asked questions

Which AI model is best at math on MathArena?

As of November 17, 2026, MathArena lists Gemini 4 Carbon (high) first at 85.4 percent plus or minus 3.0, ahead of Claude Fable 5.2 (max) at 82.8 percent and GPT-6.1 Sol (max) at 80.6 percent.

How much does Gemini 4 Carbon cost on MathArena?

MathArena's table lists an expected cost of $15.70 per problem for Carbon, against $40.20 for Fable 5.2 and $1.07 for GPT-6.1 Sol. The Decoder describes Carbon as roughly 15 times Sol's cost.

Is Carbon's lead statistically significant?

Not clearly. Its 4.8 point lead over Sol sits against intervals of plus or minus 3.0 and 2.7, and its lead over Fable 5.2 is smaller than Fable's plus or minus 5.3.

What is the best open model on MathArena?

Kimi K3.1 Pro (max) from Moonshot AI, released October 23, 2026, at 54.3 percent plus or minus 5.8 and $3.70 per problem.

What are MathArena's expected performance scores?

MathArena says they are estimates that combine observed results with item-response modeling for questions a model has not answered, reported with confidence intervals.

Why does Carbon cost so much to run?

The Decoder reports that Carbon almost never gives up, that its most common failure is exceeding the time limit, and that one run executed a script for over eleven hours without checking progress.

Sources

What each one is, and whose it is.

  1. 1

    MathArena leaderboard, MathArena (November 17, 2026)

    BenchmarkIndependent of the vendor
  2. Press reportIndependent of the vendor
  3. 3

    Gemini 4 Carbon launch post, Google (November 11, 2026)

    Vendor announcement