OpenAI GPT-6 Astra achieves 70 percent on theorem proving benchmark
A new research paper shows OpenAI's GPT-6 Astra completed formal proofs for 70 percent of tested theorems, but failed to solve any open mathematical conjectures.
FTMQ SI, written by our newsroom0 views
OpenAI's GPT-6 Astra produced a complete, machine-checked Lean proof in 70.0 percent of attempts on 20 research theorems in theoretical computer science, FourWeekMBA and Crypto Briefing reported. The findings come from the TCSExam benchmark paper posted to arXiv on November 5 by researchers from Tsinghua University, Moonshot AI, and the Shanghai Qizhi Institute. However, the same model resolved none of the 30 open conjectures included in the benchmark. [1][2]
Moonshot AI's Kimi K3.1 Pro model placed second in the evaluation, achieving a 38.3 percent success rate on theorem proving, according to FourWeekMBA. Six other artificial intelligence agents were evaluated on the benchmark suite. [1]
In formal logic and mathematics, a theorem is a statement that has been proven using deductive rules from established axioms, while a conjecture is a proposition offered without proof. Large language models have made growing progress since the mid-2020s in generating research-level formal proofs, primarily using models developed by OpenAI and Anthropic. [3][4][5]
In short
- GPT-6 Astra generated complete Lean proofs for 70.0 percent of 20 research theorems.
- GPT-6 Astra solved zero out of 30 open mathematical conjectures.
- Kimi K3.1 Pro placed second on the theorem-proving benchmark with a 38.3 percent success rate.
- The TCSExam paper was authored by researchers from Tsinghua University, Moonshot AI, and the Shanghai Qizhi Institute.
Sources
Every paragraph above points to the numbered items it rests on. Read the originals here.
- [1]GPT-6 Astra Proves 70% of Theorems but 0 of 30 ConjecturesFourWeekMBA, 9d ago (the report this story comes from)
- [2]GPT-6 Astra proves 70% of Lean theorems but solves none of 30 open conjecturesCrypto Briefing, 9d ago
Background
- [3]List of mathematical discoveries made with artificial intelligence on Wikipedia
- [4]Theorem on Wikipedia
- [5]Conjecture on Wikipedia
Our newsroom writes these reports with the help of software, from the 5 sources listed and nothing else, and checks them against those sources. Facts can still be wrong or move on; the originals are the record. Spotted a mistake? Write to daniel@monsterkong.com.
Related from FTMQ SI
Earlier reports of ours on the same people and subjects.
- Three AI labs report formal Lean proofs in scientific benchmark evaluationIndustry Labs, 9d ago
- OpenAI releases 722 manuscripts solving open math problems with frontier modelTop Stories, 42d ago
- OpenAI releases GPT-6.2 Astra and Sol modelsIndustry Labs, 4h ago
- AI agents fail to pass Department of Energy scientific paper benchmarkResearch Papers, 8d ago
- Cmpnd releases dspy-wsp for weight space programming in DSPyModels & Products, 2d ago
More in Research Papers
- Google DeepMind Releases Agentic Cheminformatics Tools for Google Antigravity8d ago
- CERN LHCb detector captures ultra rare B meson decay evidence8d ago
- AI agents fail to pass Department of Energy scientific paper benchmark8d ago
- Google releases Gemini Omni 1.2 Flash video AI model10d ago
- AI agents struggle on end to end physics research benchmark10d ago
- Google DeepMind releases Gemini 4 Carbon frontier model16d ago
Get the day in one email
Reports like this one, the top news of the last 24 hours, every morning. Free, one email a day; readers can comment under every report.
By signing up you agree to our terms and privacy policy. Unsubscribe any time.

Comments
Loading
Join the conversation
Comments are open to readers of our daily email: the top news of the last 24 hours, every morning, free. Sign up and the comment box opens.
Already on the list? Enter the same address and we will send a sign-in link.
By signing up you agree to our terms and privacy policy. Unsubscribe any time.