Google announced that its Gemini 4 Carbon model achieved the highest score on the MathArena math benchmark on November 10, 2026. The test measures how well large language models solve math problems, from basic algebra to advanced physics equations. Gemini 4 Carbon posted an overall accuracy of 85.4 % with a margin of error of ±3.0 %. It pushed OpenAI’s GPT‑6.1 Sol, which had led the chart with 80.6 % accuracy, down to third place. Anthropic’s Claude‑Fable 5.2 came in second with 82.8 % accuracy. The ranking is based on a “high” reasoning setting that lets the model take extra steps to reach a solution.
How does Gemini 4 Carbon differ from the competition?
The benchmark notes that Gemini 4 Carbon rarely gives up on a problem. In one test it ran a script for more than eleven hours without stopping, exceeding the time limit set by the evaluation. This persistence can be useful for very tough questions, but it also means the model can waste compute resources. By contrast, GPT‑6.1 Sol finished most tasks within the allotted time but sometimes abandoned a problem early. The extra compute time translates into higher cost. With the same reasoning settings, Gemini 4 Carbon costs roughly $15.70 per 1,000 tokens, while Claude‑Fable 5.2 costs about $40.20, and GPT‑6.1 Sol costs only $1.07. The price gap is about fifteen times between Gemini 4 Carbon and GPT‑6.1 Sol.
What does this mean for users?
Google’s model is currently only available through its paid API, so developers who need the strongest math performance can choose Gemini 4 Carbon if budget permits. For most everyday applications—such as tutoring apps, spreadsheet assistants, or simple homework helpers—the cheaper GPT‑6.1 Sol may be sufficient. The high cost also raises questions about scalability for large‑scale services that would need to run thousands of queries per day. Until Google releases pricing tiers or cheaper inference options, the trade‑off between accuracy and expense will guide adoption.
Why it matters
For developers building math‑heavy AI tools, Gemini 4 Carbon offers the best accuracy currently available, but the high price may limit its use to niche, high‑value applications. Most users will likely stick with cheaper models like GPT‑6.1 Sol unless they need the extra precision for complex scientific problems.
