Through the first half of 2026 AI systems moved from helping researchers to producing results researchers then verified and published. The clearest marker is Google DeepMind’s AI Co-Scientist, a Gemini-based multi-agent system whose work was peer-reviewed in Nature on 19 May 2026. Around it, reasoning models from DeepMind and OpenAI produced contributions to long-standing open problems in mathematics. These are early, bounded results, but they are a genuine shift in what “AI for science” means.

What happened

DeepMind’s AI Co-Scientist is a multi-agent system built on Gemini that generates, debates, and refines scientific hypotheses. In the Nature paper, it produced experimentally validated leads across independent labs, including in liver-fibrosis drug repurposing and other biomedical areas. Because it was peer-reviewed, it is the most solid of the year’s AI-for-science claims.

On the mathematics side, DeepMind reported on 11 February 2026 that its Gemini Deep Think reasoning mode had solved several open questions from the Erdős problems database and contributed to problems in computer science and physics under expert guidance. Separately, a preprint dated 20 May 2026, co-authored by nine leading mathematicians, described a counterexample generated by an OpenAI reasoning model that disproves the Erdős unit-distance conjecture, a decades-old problem in discrete geometry. The math results are preprints and lab reports, not yet peer-reviewed, and the exact model behind the OpenAI result is not named.

Why it matters for builders

The pattern underneath all three is the same: a reasoning model , often wrapped in a multi-agent loop that proposes, critiques, and refines, applied to a problem with a checkable answer. That is exactly the shape of a good agentic system, and it is why these results appeared in science and math first: both fields let you verify a candidate answer, so the AI can search and the human can check.

For builders, the transferable lesson is not “AI does science now.” It is that agentic systems work best where verification is cheap: math proofs, code that compiles and passes tests, hypotheses a lab can run. If your problem has a checkable answer, a propose-and-verify agent loop is worth trying. If it does not, you inherit all the reliability problems these results carefully worked around. Treat the headline claims as bounded and mostly not-yet-peer-reviewed, except the Nature-published Co-Scientist work.

Sources

Further reading