Google’s Gemini 4 Carbon has taken the top position on MathArena, becoming the first non-OpenAI model in some time to lead the mathematics benchmark. The model achieved an estimated performance score of 85.4%, moving ahead of OpenAI’s GPT-6.1 Sol and Anthropic’s Claude Fable 5.2. The result represents an important development in the competition between leading AI companies to build models capable of solving increasingly difficult mathematical problems.
The achievement was highlighted by MathArena researcher Jasper Dekoninck on November 13, 2026, after the benchmark added evaluation results for Gemini 4 Carbon and Claude Fable 5.2. However, the performance improvement comes with a significant cost difference. MathArena estimates that Gemini 4 Carbon costs approximately $15.70 per problem, compared with around $1.07 for GPT-6.1 Sol. The results demonstrate Google’s progress in advanced mathematical reasoning while raising questions about the economics of using high-end AI models.
Gemini 4 Carbon Takes First Place on MathArena
According to MathArena’s leaderboard, Gemini 4 Carbon (high) achieved an expected performance score of 85.4%, placing it ahead of several leading models from OpenAI and Anthropic.
The result is particularly notable because OpenAI models have occupied the top positions on the benchmark for an extended period.
The latest leaderboard places Anthropic’s Claude Fable 5.2 in second position, followed by OpenAI’s GPT-6.1 Sol.
| Rank | AI model | Developer | Expected performance | Estimated cost per problem |
|---|---|---|---|---|
| 1 | Gemini 4 Carbon (high) | 85.4% | $15.70 | |
| 2 | Claude Fable 5.2 (max) | Anthropic | 82.8% | $40.20 |
| 3 | GPT-6.1 Sol (max) | OpenAI | 80.6% | $1.07 |
| 4 | GPT-6 Astra | OpenAI | 78.0% | $4.03 |
| 5 | GPT-6 Sol | OpenAI | 75.9% | $4.07 |
The figures represent MathArena’s expected performance estimates across its active competitions, rather than results from a single mathematics examination.
Gemini 4 Carbon’s lead over Claude Fable 5.2 is 2.6 percentage points, while its advantage over GPT-6.1 Sol is 4.8 percentage points.
The leaderboard therefore places Google’s model at the top, although the difference between the leading systems should be considered alongside their evaluation costs and testing limitations.
What Is the MathArena Benchmark?
MathArena is an independent evaluation platform designed to measure how effectively AI models solve challenging mathematical problems.
Unlike conventional benchmarks that may contain widely circulated questions, MathArena uses problems from recent mathematics competitions, olympiads and research challenges.
Its goal is to evaluate models on questions that were published after their training periods, reducing the possibility that an AI system has simply memorised the answers.
The platform includes several competitions and evaluation categories, including:
- ArXivMath: Tests mathematical problem-solving using research-oriented questions.
- BrokenArXiv: Evaluates the ability to solve difficult mathematical problems derived from research material.
- ArXivLean: Focuses on mathematical reasoning and formal proof-related tasks.
- LeanLean: Evaluates mathematical reasoning involving the Lean proof assistant.
- Project Euler: Tests mathematical and computational problem-solving.
MathArena evaluates models across multiple questions and runs each model four times on each problem. Its overall performance score combines results across active competitions, giving individual questions equal weight.
This approach is intended to provide a broader picture of mathematical reasoning rather than relying on one isolated test.
Gemini 4 Carbon Scores 85.4% but Costs 15 Times More
One of the most important details in the latest leaderboard is the difference in operating costs.
Gemini 4 Carbon delivers the highest expected performance, but its estimated cost of $15.70 per problem is substantially higher than GPT-6.1 Sol’s $1.07.
That makes Google’s model approximately 14.7 times more expensive under the tested configurations.
Anthropic’s Claude Fable 5.2 is even more expensive, with an estimated cost of $40.20 per problem.
| Model | Expected performance | Cost per problem |
|---|---|---|
| Gemini 4 Carbon | 85.4% | $15.70 |
| Claude Fable 5.2 | 82.8% | $40.20 |
| GPT-6.1 Sol | 80.6% | $1.07 |
The cost difference matters for developers and research organisations that need to solve thousands of mathematical problems.
For an individual researcher working on a particularly difficult problem, paying more for a potentially stronger model may be reasonable. However, businesses operating AI systems at scale must consider whether the additional accuracy justifies the higher expenditure.
The results suggest that mathematical reasoning performance and cost efficiency remain separate areas of competition among AI developers.
Google’s Gemini 4 Carbon Sometimes Struggles With Time Limits
Despite topping the leaderboard, Gemini 4 Carbon has some limitations.
Dekoninck identified exceeding time limits as one of the model’s most common failure modes during testing.
The model reportedly tends to continue working on difficult problems rather than abandoning an approach when it becomes unproductive.
In one example shared by the researcher, Gemini 4 Carbon launched a script that ran for more than 11 hours without checking its progress.
This behaviour highlights a challenge with advanced reasoning models: spending more time on a problem does not always guarantee a correct answer.
Long-running computations can also increase operating costs and make the model less practical for applications requiring predictable response times.
The benchmark’s results therefore reflect not only mathematical reasoning but also how models manage tools, computation and extended problem-solving attempts.
For developers, the ability to recognise when an approach is failing can be just as important as the ability to solve difficult equations.
Claude Fable 5.2 Takes Second Place
Anthropic’s Claude Fable 5.2 secured second place with an expected performance score of 82.8%.
The result puts it ahead of OpenAI’s GPT-6.1 Sol, although it comes with the highest estimated cost among the three leading models.
MathArena estimates that Fable 5.2 costs approximately $40.20 per problem under the tested maximum-reasoning configuration.
Dekoninck also identified reasoning-extraction errors as a recurring issue during testing. These errors occur when the evaluation process encounters problems extracting or handling the model’s reasoning output.
The researcher said the team treated these errors as failures rather than ignoring them, since they affect the reliability of the evaluation process.
Fable 5.2’s second-place result nevertheless demonstrates Anthropic’s progress in mathematical reasoning and complex problem-solving.
What the Results Mean for OpenAI
OpenAI’s GPT-6.1 Sol previously held the top position on MathArena before Gemini 4 Carbon moved ahead.
Its latest expected performance score stands at 80.6%, placing it 4.8 percentage points behind Google’s model.
However, GPT-6.1 Sol has a considerable cost advantage.
At approximately $1.07 per problem, it is much cheaper to operate under the benchmark’s testing conditions than either Gemini 4 Carbon or Claude Fable 5.2.
This creates a different competitive picture. Google leads in the overall expected performance estimate, while OpenAI’s model offers a substantially lower estimated cost.
The results do not mean OpenAI has lost its broader position in AI research or that Gemini 4 Carbon is superior across every category. MathArena measures a specific set of mathematical and reasoning capabilities, and performance can differ considerably across coding, writing, research and other workloads.
The Bigger Picture
Gemini 4 Carbon’s MathArena result highlights how quickly competition in advanced AI reasoning is evolving. Google has demonstrated that its latest model can outperform leading OpenAI and Anthropic systems on a benchmark designed around challenging and relatively recent mathematical problems.
However, the cost difference reveals another important challenge for the industry. The most capable model is not necessarily the most economical choice for every application. Businesses and researchers must balance accuracy, computing expenses, response times and reliability when selecting AI systems.
The results also demonstrate why independent benchmarks matter. Instead of relying exclusively on performance claims made by AI developers, platforms such as MathArena provide a common evaluation framework for comparing different models under similar conditions.
Looking Ahead
Google’s Gemini 4 Carbon has established itself as the current MathArena leader, but maintaining that position will depend on how competing models improve and how the benchmark evolves. OpenAI and Anthropic are continuing to develop increasingly capable reasoning systems, and future releases could change the leaderboard again. Independent testing will remain important for determining whether performance improvements translate into consistent results across different mathematical challenges.
The larger question is whether Google can make Gemini 4 Carbon’s advanced reasoning capabilities more affordable and efficient. Its current performance comes at a substantially higher cost than GPT-6.1 Sol, creating a trade-off between accuracy and economics. As AI models become more capable, the competition will increasingly involve not just solving the hardest problems but doing so reliably, quickly and at a price that developers and businesses can justify.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.



