Chat with the newest AI models and work with them on your hardest problems
September 20, 2026 ·
You can now attach files to a chat.
PDFs, images, and text files such as
.tex or .csv
are read by the models and stay with the conversation.
September 17, 2026 ·
All 100 problems of the
Leipzig Benchmark
were re-run against the newest models.
Four problems were solved by none of the four leading models.
September 17, 2026 ·
Claude Opus 5, GPT-6 Astra,
GPT-6 Astra Pro,
and Claude Fable 5.1 are now available.
Every user gets free monthly credit. Prepaid credit is available individually or as a shared pool for research groups and teams.
See My Chat or contact us for shared pools at contact@sciencebench.ai.
Newest Available Frontier Models:
GPT-6 Astra
∅ per turn: 0.23cp
Claude Opus 5
∅ per turn: ?
Grok-4.3
∅ per turn: 0.56cp
Gemini 3.1 Pro
∅ per turn: 0.51cp
Premium Models (only available for grant users):
GPT-6 Astra Pro
∅ per turn: ?
Claude Fable 5.1
∅ per turn: ?
327
researchers signed up
325
benchmark problems by 102 users
4512
chat turns by 160 users
See how the newest models perform on research-level mathematics.
Public BenchmarksBrowse research-level problems contributed by the community.
Sample Problems