Google’s Gemini 4 Carbon has taken the top position on the MathArena benchmark, becoming the first non-OpenAI model in some time to lead the public ranking. Anthropic’s Claude Fable 5.2 placed second, pushing GPT-6.1 Sol below both systems.
The result comes with important operational limits. Carbon’s most common failure was exceeding the benchmark’s time limit, and the model rarely abandoned a task even when progress was slow. In one test, it started a script and let it run for more than 11 hours without checking the result. Persistence can help on difficult mathematics, but it also increases latency and compute use.
Cost separates the leaders further. Under the reasoning settings used by MathArena, Gemini 4 Carbon cost roughly 15 times as much as GPT-6.1 Sol, while Fable 5.2 cost about 40 times as much. Those ratios apply to this benchmark configuration rather than every workload.
The ranking shows that benchmark leadership and practical efficiency can point in different directions. Teams choosing a model for mathematical work need to weigh answer quality against time limits, failure behavior and total task cost. OpenAI has also discussed stronger unreleased systems, but models that are not publicly available cannot yet be compared through ordinary user access.