Researchers from Tsinghua University and Moonshot AI have introduced TCSExam, a new benchmark designed to evaluate frontier AI agents in theoretical computer science. The test consists of 100 tasks divided into four tracks, including theorem proving in the Lean language, solver engineering, construction building, and attempts to solve open conjectures.
In the primary evaluation, OpenAI’s GPT-6 Astra successfully proved 70% of the theorems presented to it. However, the model failed to resolve any of the 30 open conjectures included in the benchmark. All submissions were verified by the Lean kernel or specialized checkers to ensure accuracy, with broken arguments receiving no credit.
The benchmark is available in two versions: TCSExam-Full with 100 tasks and TCSExam-Lite with 20 tasks. In the full assessment, Moonshot AI’s Kimi K3.1 Pro achieved a score of 57.4, successfully proving 38.3% of the theorems. In the Lite version, Anthropic’s Claude Opus 5.5 outperformed Astra with a score of 72.3 compared to Astra’s 67.0.
The benchmark paper was submitted to the arXiv preprint server in early November 2026. The development of GPT-6 Astra follows OpenAI’s August 2026 release of ten machine-checkable proofs generated by an internal model, which was subsequently made public.
