Microsoft MAI-DxO diagnosis paper in Science
Result confirmed
Status
- Claim
Result confirmed
- Our reporting
- High confidence
- Verification
- Peer-reviewed
- Importance
- 3 of 5
- Last verified
- 23 October 2026
Your AI and this story
- GPT-6 Astra175 days after its cutoff
- Claude Opus 5.5114 days after its cutoff
- Gemini 4 Argon205 days after its cutoff
- Grok 4.7144 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 114 days before it.
Key facts
- Paper: ‘Sequential diagnosis with language models’, Science, Oct 22, 2026 (DOI 10.1126/science.aeh7930)
- Accuracy on complex NEJM cases (TechCrunch): MAI-DxO 83%; primary-care doctors and general internists working alone with reference materials 36%. Specialists and physician teams, who normally handle such cases, were not compared
- Model-agnostic: tested with OpenAI, Google, Anthropic, Meta and DeepSeek models; GPT-5 alone 78.9% at ~$8,400 per case, with DxO ~4 points higher and ~$3,000 cheaper; GPT-5.5 72% → nearly 87% at slightly lower cost; Gemini 2.5 Flash 46.4% → ~69%
- Limits: on emergency-department cases DxO added little (GPT-5.5 ~85% with or without it); most of the gain comes from the base models, and gains shrink as models near the ceiling; all tests are simulations, not real patients
- Next steps (Suleyman): a prospective study with Beth Israel Deaconess Medical Center (consenting patients also consult DxO; its recommendations compared with doctors’); a healthcare foundation model with Mayo Clinic due next year; he predicts a breakthrough in forecasting how medical records unfold by 2027
What happened
Science published Microsoft AI’s diagnostic-orchestrator paper on Oct 22. The system turns a language model into a virtual panel that asks questions, orders tests with costs attached and argues toward a diagnosis on cases from the New England Journal of Medicine. Microsoft AI CEO Mustafa Suleyman used the publication to restate the goal of “medical superintelligence”.
Why it matters
The result is now peer-reviewed, and it is model-agnostic: the orchestrator lifts every frontier model tested. The paper’s own data also shows its limits: most of the gain comes from better base models, it adds little on emergency cases, and the doctor comparison excludes specialists. The Beth Israel study will be the first test on real patients.
Sources
3 sources from 3 sites. Numbers match the chips in the text.
3 sources: 1 primary, 2 press
Primary
- Science: Sequential diagnosis with language models (Oct 22, 2026)science.org, paper
Press
Changes
- Filed (TechCrunch, The Verge)