Faster decisions inside apps
OpenAI’s Decisions API is now generally available. It gives developers a focused way to classify an input, route a request, or choose an agent’s next action, returning a probability, predefined choice, or score from text, images, or structured data.
OpenAI says the API can make these bounded decisions up to 10 times faster than GPT-6 Luna through the Responses API. Since beta, it has gained prompt caching, native JSON input, image instructions, U.S. and EU inference residency, and support for up to 400 questions in one request.
Testing diagnosis one step at a time
Microsoft’s new medical AI research evaluates diagnosis as a sequence of questions and tests rather than a single answer. Its SDBench uses 304 complex cases from The New England Journal of Medicine and a second set of more typical emergency-department cases.
Microsoft reports that its MAI-DxO system reached 83 percent accuracy on the complex cases and 80 percent on the emergency cases, compared with 36 percent and 47 percent for physicians in the respective evaluations. Selected benchmark cases and reference code are open, while a planned prospective study at Beth Israel Deaconess Medical Center will evaluate the system through independent physician assessments.
Also
- Explore changes across the planet Ai2 released OlmoEarth v1.3 with annual satellite embeddings and a browser tool for comparing years, finding similar places, or training a small map classifier without an account or GPU.
- Build charts people can manipulate Microsoft Research’s Flint-interactive lets agents create charts with clicking, filtering, zooming, highlighting, and connections to tables, text, or application state.
- Let an agent propose the cleanup first Arol Wright describes asking an agent to inspect a cluttered folder and propose an organization plan before moving files, with deletion and other irreversible choices left for human review.
Pass it on