Gemini 4 Carbon tops MathArena at a much higher cost
Google’s Gemini 4 Carbon has taken first place in the MathArena mathematics benchmark, becoming the first non-OpenAI model in some time to lead the ranking. It displaced OpenAI’s GPT-6.1 Sol, while Anthropic’s Claude-Fable-5.2 moved into second place. OpenAI’s models are not publicly available yet, although they have reportedly handled demanding problems such as the Navier-Stokes equations.
In testing, Gemini 4 Carbon achieved an expected performance of 85.4% ±3.0% with the high reasoning setting, compared with 82.8% ±5.3% for Claude-Fable-5.2 and 80.6% ±2.7% for GPT-6.1 Sol. The best open model listed was Moonshot AI’s Kimi K3.1 Pro, at 54.3% ±5.8%.
Gemini’s most frequent failure was running past the time limit. It rarely abandoned tasks; in one reported case, it started a script that continued for more than eleven hours without checking its progress. The tested configuration cost $15.70 ±$1.33, roughly 15 times GPT-6.1 Sol’s $1.07 ±$0.018. Claude-Fable-5.2 cost $40.20 ±$2.59, about 40 times as much.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.