Papers, benchmarks, and findings from labs and universities.
On difficult New England Journal of Medicine cases, MAI-DxO was 83.2% accurate. Generalist physicians with access to reference materials reached 36.1%.
The agents worked in parallel on existing CRISPR data from millions of mouse brain cells to produce hypotheses for scientists to test.
Across four models and three clinical tasks, the underlying model mattered more than how the agents were arranged.
E(3)-Pose, developed with Boston Children's Hospital and Inria, adjusts scans on the fly and could spare families repeat hospital visits.
Over 27 hours, the agents divided up the work themselves, kept a shared memory and killed the Ender Dragon. The run cost $1,620.
Anthropic says the 25-year-old origin of the ECDSA curve seeds makes a deliberate backdoor less likely, and it is donating the $36,864 bounty.
Researcher Matilda Glynn-Henley warns of "agentic flooding" if autonomous assistants are deployed at scale.
GPT-5.5 flagged a planted negative result in 2 of 200 reports, but in 190 of 200 once the prompt said "Be honest in your response."
Experts quoted by Scientific American say the 166-page Lean proof does not follow the approach OpenAI credited.
Anthropic says it cannot veto METR's published findings, though it can require redactions in narrow, pre-defined categories.
Thirty-five researchers built TheoPhysEval from routine research calculations that would take a subfield expert about a day or less to solve.
Haydn Belfield says the share of code changes handed fully to AI rose from 7.5 to 11.9 percent, and every change still gets human review.