News
November 15, 2026 · The Decoder

Gemini 4 Carbon tops MathArena's leaderboard, though at a far higher listed cost

MathArena's newest results put Google's Gemini 4 Carbon first overall, ahead of Claude Fable 5.2 and GPT-6.1 Sol, but its estimated cost is many times that of OpenAI's model.

MathArena, a math benchmark for AI models run by ETH Zurich's SRI Lab and INSAIT, added results for Google's Gemini 4 Carbon and Anthropic's Claude Fable 5.2 on November 13, 2026. According to The Decoder, which reported the figures on November 14, Gemini 4 Carbon now holds the top overall spot, pushing OpenAI's GPT-6.1 Sol down the table.

What's new

The Decoder lists MathArena's overall recommendations as follows:

  • 1st, Gemini 4 Carbon (high reasoning): expected performance of 85.4%, with an expected cost of $15.70.
  • 2nd, Claude Fable 5.2 (max): 82.8%, at $40.20.
  • 3rd, GPT-6.1 Sol (max): 80.6%, at $1.07.
  • Best open model, Kimi K3.1 Pro (max) from Moonshot AI: 54.3%, at $3.70.

By The Decoder's estimate, running GPT-6.1 Sol is about 15 times cheaper than Carbon and about 40 times cheaper than Fable 5.2 at the reasoning settings used. It also reports that Carbon most often fails by running past the time limit and rarely gives up on a problem; according to the outlet, one script it launched kept going for over eleven hours while the model never looked in on it.

The Decoder notes that OpenAI models it has reported on for advanced equation work are not yet publicly available, so the table covers released models only.

Why it matters

Analysis: for people choosing a model for math-heavy work, the table shows the top three are close in accuracy but far apart in price. A few points of benchmark score may not justify a cost many times higher for routine tasks, while the hardest problems may warrant it. Benchmarks also measure a narrow skill, so results on your own workload can differ. Gemini and Claude users who rely on high reasoning settings may want to watch their usage costs.

Source: The Decoder