Chat with the newest AI models to work on your hardest problems

Discuss your mathematics with GPT, Claude, Gemini and others in parallel.

Free 30-day rolling credit for every researcher. Your chats are not used for AI training. 370 researchers signed up.

The following frontier models are currently available

ModelAll models are queried through the business APIs of their companies. Your chats are not used for AI training. We have set up the service so that the companies store your data for as short a time as possible. Research problems solvedCorrect answers on the 100 research-level problems of the Leipzig Benchmark. The settings of each model are listed there. Answers kept by usersHow often users continued the conversation with an answer of this model family (all versions of the model, such as all Claude Opus versions). Median answer timeThe median time a model takes for an answer in the chat, at its default thinking effort. CostThe relative cost of a median answer in the chat, at the default thinking effort. $$$$ marks the most expensive model available.
Claude Opus 5.5
85%
70% 1m07s $$$$
Gemini 4 Neon 54% 26s $$$$
GPT-6.2 Astra 96% 1m31s $$$$
GPT-6.2 Sol 85% 1m00s $$$$
GPT-6.2 Astra Pro premium model for prepaid credit 95% 5m25s $$$$
Claude Fable 5.2 premium model for prepaid credit 74% 3m30s $$$$

Based on 100 benchmark problems and 1945 chats with 5698 chat turns and 5576 model answers.

November 21, 2026 · The chat now runs the newest models from Anthropic, OpenAI and Google: Claude Fable 5.2, GPT-6.2 Astra, GPT-6.2 Astra Pro, GPT-6.2 Sol, and Gemini 4 Neon.
November 19, 2026 · For your students: MathTutor, our sibling project, is an AI tutor that discusses exercises and the students' own solution attempts, giving hints instead of finished solutions (Bastani et al., PNAS 2025). MathTutor is a pilot project at Ruhr University Bochum.
October 6, 2026 · Version 2 of our paper Benchmarks in Leipzig is now on arXiv. We believe that the current way of submitting benchmark problems has reached its limits, and we are looking into new ways of benchmarking. In the meantime, we are closing the submission system.