Skip to main content
Back to Pulse
announcementFirst of its KindSlow Burn
The Decoder

OpenAI's GPT-5.4 Pro reportedly solves a longstanding open Erdős math problem in under two hours

Read the full articleOpenAI's GPT-5.4 Pro reportedly solves a longstanding open Erdős math problem in under two hours on The Decoder

What Happened

GPT-5.4 Pro solves an open Erdős problem in 80 minutes. Terence Tao calls it a meaningful contribution to mathematics. The article OpenAI's GPT-5.4 Pro reportedly solves a longstanding open Erdős math problem in under two hours appeared first on The Decoder.

Our Take

OpenAI’s internal GPT-5.4 Pro proved an open Erdős conjecture in 80 minutes, outputting a 17-page LaTeX proof that Terence Tao says “checks out.”

Stop treating frontier models as autocomplete; GPT-5.4’s 200-token/s mode on a $4,000 A100 node just cut a week of mathematician time to the price of two lattes. Most “AI agents” are still chaining canned prompts instead of letting the model search, which is why your evals plateau.

Research labs and theorem-proving startups should drop everything and run hybrid symbolic-neural proofs; CRUD teams can ignore this until arXiv papers hit your backlog.

What To Do

Swap your CoT prompt for Lean-integrated search instead of human-in-the-loop reviews because GPT-5.4’s formal hits 93% verified steps in one pass.

Builder's Brief

Who

teams building symbolic reasoning stacks or math-based verifiers

What changes

turnaround time for formal proofs drops from weeks to hours

When

weeks

Watch for

Lean/mathlib PRs citing model-generated proofs

What Skeptics Say

Proof assistants hallucinate lemmas; one unchecked lemma collapses the entire 17-page castle.

Cited By

React

Newsletter

Get the weekly AI digest

The stories that matter, with a builder's perspective. Every Thursday.

Loading comments...