GPT-6 Astra just solved 5 math problems humans couldn't crack. Every other model scored zero

GPT-6 Astra just solved 5 math problems humans couldn't crack. Every other model scored zero

Here's a sentence that should make every mathematician pause: an AI just solved problems that humans haven't been able to crack for decades. And it's the only AI that could.

On September 4, 2026 (UTC+8), researchers from Epoch AI and the University of Manchester released the FrontierMath Erdős (FME) benchmark — a collection of 68 unsolved mathematical problems named after the legendary Hungarian mathematician Paul Erdős, who famously offered cash prizes for solutions to his favorite open questions. They threw five of the world's most advanced AI models at it. Four of them scored exactly zero. GPT-6 Astra solved five.

What the benchmark actually tests

The FME isn't your standard math test. These aren't textbook problems with known answers. They're genuine open problems — questions that have stumped professional mathematicians, some for over half a century. Solving even one would be a publishable result. Solving five is unheard of for an AI.

The five models tested were the current frontier: GPT-6 Astra, GPT-5.6 Sol, and three other top-tier models (including Anthropic's Fable 5.1, which had just appeared in AWS Bedrock's API routing). The results were stark:

Sources: Epoch AI / University of Manchester, FrontierMath Erdős benchmark (September 2026). FME uses a weighted scoring methodology, so the 3% score reflects partial credit and proof quality across the 68 problems.

A 3% score might sound modest. But when the baseline is zero across every other model on Earth, "5 out of 68" isn't a small gap — it's a different category of capability.

This isn't Astra's first math flex

The Erdős results come on top of Astra's already-impressive official math benchmarks. OpenAI's own technical documentation shows Astra scoring 97.6% on FrontierMath Tier 4 — a research-level math test that previous models struggled to pass at all.

More strikingly, Astra has been credited with advancing actual mathematical knowledge:

• Prime pair distance: Astra pushed the upper bound on prime pair distances from 246 down to 186 — a result that had stood for years
• Large prime gaps: Astra improved a bound on large prime gaps that had remained unchanged for 80 years

These aren't benchmark scores. These are contributions to mathematics. The kind of thing that normally requires a PhD and years of focused work.

Why this matters beyond math

The Erdős benchmark is significant because it measures something different from standard tests. MMLU, GSM8K, and HumanEval all have known answers — they test whether a model can reproduce human knowledge. FME tests whether a model can produce new knowledge.

That's a much higher bar. It's the difference between a student who aces exams and a researcher who publishes original work. And it's exactly the kind of capability that people point to when they argue about whether "AGI" has arrived.

OpenAI president Greg Brockman's line at the Astra launch — "Welcome to the AGI era" — drew plenty of skepticism. But when your model is the only one on Earth that can solve unsolved math problems, the claim starts to carry weight.

The caveats

Before anyone declares the problem of mathematics solved, some honesty:

• 5 out of 68 is still 63 unsolved. The vast majority of Erdős problems remain beyond even Astra's reach.
• "Solved" needs verification. The FME benchmark identifies candidate solutions, but each one requires independent verification by human mathematicians. A wrong answer that looks plausible is worse than no answer.
• Other models may catch up. GPT-5.6 Sol scored zero today, but the gap between model generations is shrinking fast.
• We don't know the full methodology yet. The Epoch AI / Manchester paper is new, and the community will need time to scrutinize how problems were selected and how solutions were verified.

What to watch

• Independent verification — mathematicians checking whether Astra's five Erdős solutions actually hold up
• The full paper — Epoch AI and Manchester's methodology, problem selection, and failure analysis
• Competitor response — whether Anthropic, Google, or others release models that can score non-zero on FME
• More open problems — whether Astra (or its successors) starts tackling problems beyond the Erdős catalog

For now, the headline is simple: on a test where every other top AI model scored zero, GPT-6 Astra solved five problems that humans couldn't. Whether that's AGI or just a very impressive math assistant is up for debate. But it's not nothing.