GPT-6 Astra proves 70% of Lean theorems but solves none of 30 open conjectures

OpenAI’s GPT-6 Astra can now prove 70% of the theorems put in front of it in a new theoretical computer science benchmark. It also solved exactly zero of the 30 open conjectures it was handed.

That split is the headline result of TCSExam, a benchmark built by researchers from Tsinghua University and Moonshot AI.

What TCSExam actually measures

TCSExam is a set of 100 tasks, each reviewed by experts, aimed at frontier AI agents working on theoretical computer science.

The tasks are split into four tracks. The first is theorem proving in Lean. The second is solver engineering. The third is building constructions that beat existing human benchmarks. The fourth is the hard one: attempts to prove open conjectures.

Lean is a formal proof language, which means a proof written in it has to satisfy a strict mechanical checker. Either the Lean kernel accepts the proof or it does not.

Every submission gets verified by the Lean kernel or by a checker built for that specific task. A model that confidently produces a plausible but broken argument scores nothing.

The benchmark comes in two sizes. TCSExam-Full contains all 100 tasks. TCSExam-Lite trims that down to 20.

The scoreboard

On the full version, GPT-6 Astra took first place with an index score of 71.9.

Most of that lead came from theorem proving. Astra completed 14 out of 20 tasks in that track, for a 70% success rate.

Second place went to Moonshot AI’s own Kimi K3.1 Pro, which posted a full-benchmark score of 57.4. Kimi proved 38.3% of the theorems it was tested on.

The Lite version told a slightly different story. On the 20-task variant, Anthropic’s Claude Opus 5.5 came out ahead, scoring 72.3 to Astra’s 67.0.

The conjecture wall

In the primary evaluation, no agent resolved any of the 30 open conjectures. The success rate across the board was 0%.

How we got here

In August 2026, OpenAI released ten machine-checkable proofs produced by an internal model called Astra. That model was later made public as GPT-6 Astra.

The benchmark paper was submitted to the preprint server arXiv in early November 2026. As a preprint, it has not necessarily gone through formal peer review.