POLICY · MODELS · #3597
Gemini 4 Carbon takes MathArena crown, displacing GPT-6.1 Sol
Google's Gemini 4 Carbon ranked first on the MathArena math benchmark, knocking GPT-6.1 Sol from the top spot; Anthropic's Claude Fable 5.2 placed second. Testers noted Carbon commonly exceeds time limits and rarely abandons tasks, and under the testing reasoning settings Carbon and Fable 5.2 cost roughly 15× and 40× more than GPT-6.1 Sol, respectively.
KEY POINTS
- Google's Gemini 4 Carbon ranked first on the MathArena math benchmark, knocking GPT-6.1 Sol from the top spot; Anthropic's Claude Fable 5.2 placed second.
- Testers noted Carbon commonly exceeds time limits and rarely abandons tasks, and under the testing reasoning settings Carbon and Fable 5.2 cost roughly 15× and 40× more than GPT-6.1 Sol, respectively.
- Indicates non-OpenAI models can lead on challenging math benchmarks while highlighting trade-offs of higher capability — longer runtimes and substantially higher inference cost.
WHY IT MATTERS
Indicates non-OpenAI models can lead on challenging math benchmarks while highlighting trade-offs of higher capability — longer runtimes and substantially higher inference cost.