GPT‑5 cracks an Erdős problem and spots a cell mystery in minutes

OpenAI just put some receipts behind the GPT‑5 hype, co-publishing a paper with heavy-hitters from Vanderbilt, UC Berkeley, Columbia, Oxford, Cambridge, and Lawrence Livermore National Lab. The report isn’t a benchmark flex—it’s a collection of case studies showing what happens when domain experts treat the model like a brilliant, sometimes erratic postdoc. The numbers are stark. One researcher, Derya Unutmaz, spent months staring at a baffling shift in human immune cells. GPT‑5 looked at an unpublished chart and, within minutes, named the likely mechanism and proposed an experiment that confirmed it. That kind of speed isn’t just convenient—it compresses the timeline from observation to treatment insight. In math, Mehtaab Sawhney and Mark Sellke were stuck on the last step of a decades-old Erdős problem. The model suggested a specific insight about how one odd number breaks the pattern, and they finished the proof. It didn’t write the paper, but it unstuck the thinking.

Then there’s the less glamorous, equally important work in optimization. Sébastien Bubeck and Christian Coester used GPT‑5 to hunt for failure modes in a common decision-making algorithm used in robotics and routing. The model didn’t just find a clear, crisp example where the method falls apart—it also improved a classic result in the field. The paper is refreshingly honest about limitations. These aren’t autonomous scientists; they’re reasoning engines that work in dialogue with people who know how to prod, critique, and steer them. A striking detail is the emphasis on skill. OpenAI describes effective use as a learned practice—knowing when to push back, how to decompose problems, and what to validate independently. It’s a far cry from the ‘ask a question, get truth’ fantasy.

The strategy here is dual-track. Specialized tools like simulation engines and protein databases handle precision work; scaled foundation models handle cross-disciplinary reasoning, literature synthesis, and conceptual leaps. The paper argues these paths reinforce each other. You don’t want a language model running a physics sim, and you don’t want a database dreaming up novel mechanisms.

What’s genuinely new is the documented breadth—biology, math, algorithms, materials science—all with named researchers and specific results. The survey numbers OpenAI leads with are brutal: 60% of Americans say breakthroughs reach them too slowly, 73% want better ways to accelerate discovery. GPT‑5 won’t fix that pipeline on its own, but showing that human-AI teams can shave months off specific, hard problems is a more compelling argument than any benchmark score. The question that lingers is whether these gains are scattered miracles or the early shape of a repeatable method.