OpenAI tests GPT-5 on hard science

Hitting a dead end on a research paper for months is a familiar kind of academic pain. OpenAI just published a collaborative paper with top researchers from Oxford, Cambridge, UC Berkeley, Columbia, and Lawrence Livermore National Laboratory, detailing how GPT-5 is being used to unblock real scientific breakthroughs.

The paper compiles early case studies across mathematics, biology, physics, and computer science. This is not about writing summaries or formatting bibliographies, but about generating proofs, spotting mechanisms, and stress-testing algorithms.

Cracking Erdős puzzles and immune cells

One of the most striking results came from pure mathematics. Researchers Mehtaab Sawhney and Mark Sellke were stuck on the final hurdle of a decades-old open problem originally proposed by Paul Erdős. GPT-5 offered a novel observation on how a specific odd number disrupted the pattern, which gave them the breakthrough needed to finish the proof.

In medicine, a team led by Derya Unutmaz had spent months trying to understand a puzzling shift in human immune cells. When presented with an unpublished chart, GPT-5 pinpointed the likely biological mechanism within minutes and recommended a wet lab experiment that ended up validating the hypothesis.

Computer science saw similar wins. Researchers Sébastien Bubeck and Christian Coester wanted to know if a common routing and robotics heuristic was actually reliable. GPT-5 produced a clean counterexample proving the method can fail, while also improving an established result in optimization theory.

Researchers also noted that conceptual literature review got a noticeable lift. Instead of basic keyword searches, the model successfully connected conceptual dots across different languages and obscure papers that researchers had missed.

Keep the human in the loop

OpenAI made sure to add a massive asterisk to these findings. Every case in the report was carefully curated, meaning it does not represent an average run or expose the full scope of failure modes.

Hallucinations remain a major obstacle. GPT-5 still fabricates citations, pitches plausible sounding biological mechanisms that are totally wrong, and misses subtle domain nuances. If an expert is not there to constantly push back and verify every step, the model easily chases dead ends.

OpenAI notes that effective use is an acquired skill. The model does not conduct autonomous science; it acts as an ultra-fast sparring partner that requires rigorous oversight.

Why it matters

We think this paper highlights an important pivot in the AI for science race. While projects like AlphaFold focused on narrow, dedicated architectures for molecular biology, OpenAI is proving that frontier general reasoning models can bridge unexpected gaps across math, code, and wet lab hypotheses.

For working scientists and advanced engineers, this is worth testing immediately as an ideation and literature partner, especially when stuck on proof steps or seeking cross-discipline analogies. However, the cost of entry is deep subject-matter expertise; novice users will simply get buried under plausible-sounding nonsense.

The real danger lies in scientific noise. Because these models generate convincing arguments so effortlessly, peer review and lab validation will face immense pressure sorting brilliant leaps from fabricated hallucinations. The tool is exceptionally sharp, but only when held by someone who already knows where the cuts should land.