GPT-5.6 reportedly cracks decades-old Erdős problems amid hype debate

Between July 14 and 15, discussion of GPT-5.6's mathematical ability heated up rapidly. Multiple testers claimed the model solved several Erdős problems in succession, and the news was framed by outlets such as Scientific American as "ChatGPT has proved another 50-year-old mathematical conjecture," pushing large-model math reasoning from competition problems and standard benchmarks toward genuine research questions. The original author, however, later said the matter had been overhyped.

Key details

The most specific first-hand account came from ThunderBeanage, who said that during a week of stealth testing he used GPT-5.6 Pro to obtain solutions to several Erdős problems and had submitted five of them (730, 671, 948, 346, 1139) to the site, with some still awaiting verification. The center of discussion was Erdős #793, a problem on the asymptotic behavior of 2-primitive sets that is over 50 years old; jdlichtman first reported the result (relayed by Sebastien Bubeck), saying GPT-5.6 produced a very short and elegant construction. Another tester continued probing the same problem with GPT-5.6 Sol Ultra and arrived at a new construction.

Spread and controversy

Scientific American and derivative coverage condensed the event into "ChatGPT has proved a 50-year-old conjecture," spreading the discussion beyond technical circles. A subsequent repost noted that the original author explicitly called the matter overhyped: the proof itself may be correct, but it should not be read as a major mathematical breakthrough.

A separate claim to keep distinct

The cluster also includes a claim relayed by 新智元: it states that OpenAI officially asserted GPT-5.6 Sol Ultra, working through 64 concurrent agents, proved the "cycle double cover conjecture" in under an hour, with the result verified by Codex. Within this batch of posts the claim has no supporting first-hand source and is not part of the same evidentiary chain as the Erdős discussion, so it should be treated separately.

2026-07-14 ~ 2026-07-15 · 10 related posts

• GPT-5.6 Solves Another 50-Year Math Problem — FlorianGallwitz · 2026-07-14
• GPT-5.6 Proposes New Construction for Erdős Problem — HZoete · 2026-07-14
• GPT-5.6 Solves Another Erdős Problem — socoolandawesome · 2026-07-14
• [source] GPT-5.6 Pro Solves 5 Erdős Problems — ThunderBeanage · 2026-07-14
• [source] ChatGPT Said to Help Prove Mathematical Conjecture — scientificamerican · 2026-07-14
• GPT-5.6 Solves Old Problem, Gets Hype Backlash — _onionesque · 2026-07-15
• [source] Rumored Math Breakthrough by GPT-5.6 — 新智元 · 2026-07-15

3 near-duplicate retellings: lukaszkaiser · scientificamerican · throwawaykJQP7kiw5Fk