Models New
Gemini 4 Carbon takes first place on the MathArena maths benchmark, ahead of Claude Fable 5.2, at about 15 times GPT-6.1 Sol's cost
Published:
MathArena, the benchmark run by ETH Zurich's SRI Lab and INSAIT on recent maths competition and research problems, added Gemini 4 Carbon and Claude Fable 5.2 on 13 November. Google's Carbon, at high reasoning, now ranks first overall at 85.4% expected performance (plus or minus 3.0 points), ahead of Anthropic's Fable 5.2 at maximum reasoning on 82.8% (plus or minus 5.3) and OpenAI's GPT-6.1 Sol at 80.6% (plus or minus 2.7).
Expected cost per problem is $15.70 for Carbon, $40.20 for Fable 5.2 and $1.07 for GPT-6.1 Sol, according to the leaderboard. The best open model listed is Moonshot AI's Kimi K3.1 Pro at 54.3%. The Decoder reports that Carbon's most common failure is exceeding the time limit, and that in one run it launched a script that ran for more than eleven hours without checking its progress.
Caveat. The three leaders' error ranges overlap, so the order is not a clear win. Scores depend on the reasoning settings and harness MathArena used, and this is one benchmark of maths ability.
Sources