Gemini 4 Carbon tops MathArena at 15 times the price of GPT-6.1 Sol
Carbon almost does not stop a task. One script ran for more than eleven hours without a check.
Google made Gemini 4 Carbon. It is at the top of MathArena, a math benchmark (a standard test for models). It is the first model that is not from OpenAI to hold this position in a long time. GPT-6.1 Sol is below it, and Claude Fable 5.2 of Anthropic is second. With the reasoning in the test, Carbon costs about 15 times as much as GPT-6.1 Sol.
What is new
Carbon is possibly one of the top public models for math. OpenAI has models that are not public at this time. Reports tell that these models found answers to hard equations, such as the Navier-Stokes equations.
The numbers
- Carbon costs about 15 times as much as GPT-6.1 Sol.
- Fable 5.2 costs about 40 times as much.
The source gives no scores for the models.
What is not known
Carbon fails most frequently because it goes above the time limit. It almost does not stop a task. One script ran for more than eleven hours and did not check its work. This can be good or bad.
Insight: The cheapest model changes with the task
On 4 November 2026, SemiAnalysis gave Carbon the GDPx test. It cost $16.29 for each task, against $51.41 for Claude Opus 5.5. On MathArena, Carbon costs about 15 times as much as GPT-6.1 Sol. A test can show a different winner on price, and one rank does not give the cost. Test 2 models on your tasks and compare the cost of each task.
Sources
Posted
