Particle.news

OpenAI Unveils Decisions API for Fast, Choice-Based Model Outputs

The limited preview signals a move toward decision-first models that promise much cheaper per-action monitoring while leaving calibration and adversarial robustness unresolved.

Overview

  • OpenAI announced the Decisions API at its Dev Day on Tuesday, offering a Luna endpoint that accepts a predefined set of options and returns a ranked choice with probabilities for each option.
  • The product appears functionally similar to TypeSafe’s Jev, which was released earlier in September and is designed to output typed choices and calibrated probabilities for software automation.
  • The Decisions API is currently a limited preview and TechCrunch has not seen broad developer use yet, so practical performance, latency, and real-world calibration remain unverified.
  • Hackathon demos using Jev showed large cost differences for per-action agent monitoring compared with a frontier LLM, but those figures are illustrative and security practitioners continue to prefer shadow testing, full-distribution logging, batching, and human fallbacks.
  • Startups, open-source projects, and incumbents are racing to ship Jev-like decision models, and the main tradeoffs for users will be price, hosting and privacy, empirical calibration, and robustness to adversarial inputs.