started · updated
Google Gemini 4 Carbon tops MathArena benchmark
Google’s Gemini 4 Carbon has achieved first place on the MathArena mathematics benchmark, surpassing OpenAI’s GPT-6.1 Sol and Anthropic’s Claude Fable 5.2. While Carbon leads in performance with a score of 85.4%, its cost per benchmark run is approximately 15 times higher than GPT-6.1 Sol.
Industry analysis indicates that the cost disparity stems from the intensity of work performed per task. Google positions Carbon for cybersecurity and long-running tasks, noting that the model is designed to persist through complex operations, sometimes running scripts for over eleven hours without yielding.
Other notable performers on the benchmark include Claude Fable 5.2 in second place and Moonshot AI’s Kimi K3.1 Pro, which is recognized as the top open model on the list.
Entities
AMD · AWS · Anthropic · Apple · Broadcom · Gemini 4 Carbon · Google · Google Cloud · IBM · MathArena · Meta · Moonshot AI