Leadership change on MathArena
The MathArena math benchmark, long led by OpenAI models, has a new leader. Google's Gemini 4 Carbon has overtaken OpenAI's GPT-6.1 Sol to take first place. Anthropic's Fable 5.2 also ranks near the top.
A big cost gap
Despite the shift at the top, the cost picture favors OpenAI. According to the report, Gemini 4 Carbon is roughly 15 times more expensive to run than GPT-6.1 Sol. Higher accuracy, in other words, comes with a much higher operating cost.
Test behavior
During evaluation, Gemini reportedly exceeded the given time limit and almost never gave up on a task. That suggests the model is persistent on hard problems, but the persistence shows up in time and cost.
What it means
Benchmarks like MathArena are used to measure a model's mathematical reasoning ability. The ranking change shows Google is a serious rival to OpenAI in reasoning-focused models, though the price-performance balance remains debatable.
