Uploaded October 2026 | Updated November 2026, 3 days ago
OpenAI released more than 700 AI-written math manuscripts. Only roughly 42% of the main results come with machine-checkable Lean proofs.
On October 6, 2026, OpenAI published a repository of more than 700 mathematics manuscripts, grouped into 372 problem families. They were produced with an internal OpenAI model that has not been released. On average, each result used about three hours of ChatGPT Pro thinking compute.
The checkable part: roughly 42% of the main results come with Lean formalizations, which are proofs a computer can verify step by step. OpenAI itself warns that results without them may contain errors.
Publication is not peer review or acceptance by mathematicians, and the repository does not claim these problems are solved and accepted. Treat it as a large research corpus to inspect, not hundreds of accepted breakthroughs.
About the count: the repository README listed 719 manuscripts when checked on October 8, while press coverage reported 722. This description says "more than 700" to stay correct either way.
My take (opinion): start with the proofs a machine has checked. For the rest, mathematicians still have homework.
Would you trust an AI-written proof more if it came with a Lean check?
Event: October 6, 2026. Research checked: October 9, 2026.
Byte’s Report — research with Byte, commentary by Dave.
Sources:
Sharing AI progress in mathematics — OpenAI — 2026-10-06
openai.com/index/sharing-ai-progress-in-mathematics
openai/math README — OpenAI (GitHub) — 2026-10-06; checked 2026-10-08
github.com/openai/math
#AI #OpenAI #Mathematics #Lean #BytesReport
OpenAI released more than 700 AI-written math manuscripts. Only roughly 42% of the main results come with machine-checkable Lean proofs.
On October 6, 2026, OpenAI published a repository of more than 700 mathematics manuscripts, grouped into 372 problem families. They were produced with an internal OpenAI model that has not been released. On average, each result used about three hours of ChatGPT Pro thinking compute.
The checkable part: roughly 42% of the main results come with Lean formalizations, which are proofs a computer can verify step by step. OpenAI itself warns that results without them may contain errors.
Publication is not peer review or acceptance by mathematicians, and the repository does not claim these problems are solved and accepted. Treat it as a large research corpus to inspect, not hundreds of accepted breakthroughs.
About the count: the repository README listed 719 manuscripts when checked on October 8, while press coverage reported 722. This description says "more than 700" to stay correct either way.
My take (opinion): start with the proofs a machine has checked. For the rest, mathematicians still have homework.
Would you trust an AI-written proof more if it came with a Lean check?
Event: October 6, 2026. Research checked: October 9, 2026.
Byte’s Report — research with Byte, commentary by Dave.
Sources:
Sharing AI progress in mathematics — OpenAI — 2026-10-06
openai.com/index/sharing-ai-progress-in-mathematics
openai/math README — OpenAI (GitHub) — 2026-10-06; checked 2026-10-08
github.com/openai/math
#AI #OpenAI #Mathematics #Lean #BytesReport










