Gemini 4 Carbon leads math tests
Gemini 4 Carbon, Google's model, topped the MathArena benchmark on November 14, 2026. But its lead comes at a high cost compared to GPT-6.1 Sol, which is much cheaper to use.
Gemini 4 Carbon, Google's model, took first place in the MathArena benchmark on November 14, 2026, according to the measurement published by Matthias Bastian. It becomes the first non-OpenAI model to dominate this test for some time, ahead of GPT-6.1 Sol and Claude Fable 5.2. The achievement counts for complex calculation uses, but it comes with a much heavier bill. And math, apparently, does not come cheap in bulk pricing.
Une première place, mais pas un ticket d’entrée bon marché
On MathArena, Gemini 4 Carbon achieves an expected score of 85.4% ±3.0%, compared to 82.8% ±5.3% for Claude-Fable-5.2 and 80.6% ±2.7% for GPT-6.1 Sol. In practice, this places it at the top of the ranking, with a real but not overwhelming lead over its closest competitors.
The downside is clear: with the reasoning settings used in the test, Gemini 4 Carbon costs about 15 times more than GPT-6.1 Sol, while Claude Fable 5.2 comes to nearly 40 times the reference cost observed for Sol. At this level, mathematical performance is paid for in computation, time, and budget.
Le prix du raisonnement, ou quand le modèle insiste un peu trop
The benchmark also highlights a particular behavior of Gemini 4 Carbon: its main failure mode is exceeding the time limit. The model rarely abandons a task, which can seem like useful persistence or a tendency to stubbornness at the wrong moment.
In one case cited by the source, it ran a script for more than eleven hours without checking its progress. This kind of behavior interests both product teams and engineers, as it directly affects operational reliability, not just the final score.
Ce que le benchmark dit du marché des modèles
The comparison is not limited to a Google-OpenAI duel. It mainly shows that a model can rise to the top on a very targeted use while becoming significantly more expensive to operate. For companies, the choice is therefore not only based on the raw quality of responses but on the balance between performance, inference cost, and execution stability.
According to the source, the cited OpenAI models are not yet public, which makes the hierarchy less simple to interpret in the long term [to be verified]. Meanwhile, Gemini 4 Carbon establishes itself as a reference for mathematical tasks, even if its leadership position looks more like a technical demonstration than an economic victory.
By the numbers
- 85.4% ±3.0% — expected score of Gemini 4 Carbon on MathArena.
- 82.8% ±5.3% — expected score of Claude-Fable-5.2 on MathArena.
- 80.6% ±2.7% — expected score of GPT-6.1 Sol on MathArena.
- 15 times — estimated extra cost of Gemini 4 Carbon compared to GPT-6.1 Sol.
- 11 hours — duration of a script run without progress check.