Gemini 4 Carbon tops MathArena but costs roughly 15 times as much as GPT-6.1 Sol

Gemini 4 Carbon tops MathArena but costs roughly 15 times as much as GPT-6.1 Sol

Google's Gemini 4 Carbon has become the first non-OpenAI model in a while to top the MathArena math benchmark, according to The Decoder. It knocked GPT-6.1 Sol from the top spot, and Claude Fable 5.2 came in second. The Decoder says this makes Carbon potentially one of the most capable publicly available models for math tasks.

The report adds a caveat about OpenAI. Its models, which recently allegedly solved demanding equations such as the Navier-Stokes equations, are not publicly available yet. The report does not say which OpenAI model that refers to, and gives no verification for the claim.

The report also describes how Gemini 4 Carbon behaves. Its most common failure mode is exceeding the time limit. The model almost never gives up on a task, a trait The Decoder says can be read as either positive or negative. In one case it launched a script that ran for more than eleven hours without checking on its progress.

The lead comes at a price. With the reasoning settings used in testing, Carbon and Fable 5.2 cost roughly 15 and 40 times as much as GPT-6.1 Sol, respectively. The report gives only these multiples, not absolute prices or scores beyond the top two places.

Key facts

  • Google's Gemini 4 Carbon is the first non-OpenAI model in a while to top the MathArena math benchmark, displacing GPT-6.1 Sol.
  • Claude Fable 5.2 placed second on MathArena.
  • With the reasoning settings used in testing, Carbon and Fable 5.2 cost roughly 15 and 40 times as much as GPT-6.1 Sol, respectively.
  • Carbon's most common failure mode is exceeding the time limit; in one case it launched a script that ran for more than eleven hours without checking on its progress.
  • OpenAI's models, which allegedly solved equations like the Navier-Stokes equations, are not publicly available yet.

Why it matters

MathArena leadership has gone to a non-OpenAI model for the first time in a while, and Gemini 4 Carbon displaced GPT-6.1 Sol to get there. The Decoder calls Carbon potentially one of the most capable publicly available models for math tasks. The win comes with a steep cost gap: roughly 15 times GPT-6.1 Sol's cost, and roughly 40 times for second-placed Claude Fable 5.2, with the reasoning settings used in testing.

Who it affects

Anyone choosing a model for math-heavy work will weigh the top ranking against the cost multiples. The result also bears on Google, OpenAI and Anthropic, whose models are named in the report. OpenAI's own models, which allegedly solved demanding equations, are not publicly available yet.

How to use it

The report offers a rough guide to the trade-off rather than a price list. Carbon ranks first on MathArena but costs roughly 15 times as much as GPT-6.1 Sol with the reasoning settings used in testing. Carbon's most common failure is exceeding the time limit, so long-running tasks are the thing to watch. No absolute prices (per token or per task) are given, only the 15x and 40x multiples, and only for the reasoning settings used in testing.

How solid is it

This is a secondary report on a benchmark result; no statement from Google or Anthropic is quoted. The ranking and the cost multiples are stated by The Decoder in its own voice. No MathArena scores, percentages or rankings beyond the top two places are given. The source does not say who ran the testing or what the reasoning settings were. The OpenAI Navier-Stokes claim is hedged with 'allegedly' and no source or verification is given for it.

Risks and caveats

The cost multiples are approximate ('roughly') and tied to the reasoning settings used in testing, so they may not hold under other settings. Carbon's persistence cuts both ways: it almost never gives up, but one script ran for more than eleven hours without the model checking on its progress, and exceeding the time limit is its most common failure. The source does not say which task the eleven-hour script belonged to, or what happened to it afterwards. No release date for Gemini 4 Carbon is given, and no date for the MathArena result.