OpenAI Decisions API: a $0.10 Endpoint That Only Answers Yes, No, or Which
OpenAI put the Decisions API into public beta on Tuesday, and Hacker News gave it 381 points. It is a new endpoint, POST /v1/decisions, that takes text or images plus a list of questions and returns typed answers: a probability that a predicate holds, a choice from a fixed set, or a score against ordered levels. No free-form generation. OpenAI says it answers about 10x faster than the Responses API. The only model available is gpt-6-luna, and the price is $0.10 per million input tokens with no output, cache-read or cache-write charges at all, because there is no output to charge for. General availability is "in the coming weeks."
This is the decision-model category going mainstream at the largest API provider. TypeSafe AI's Jev started the thread. AWS followed with Strands Decider 2B and Cloudflare with Clef, both open weights. Musubi shipped PolicyLM for moderation on Monday. Now OpenAI has the same shape as a hosted product: pick from options, return a calibrated number, do it in one forward pass, make it cheap enough to put in front of every request. The three question types map directly onto what agent harnesses actually need. Predicate for "did this step succeed." Choice for routing. Score for severity or quality gating. Each answer carries a confidence, and refusals come back as a typed answer rather than an error.
The timing with the rest of the week is not an accident. Docker Agent's v1.149.0 release the same day added "OpenAI Decisions as a native evaluator backend," so one harness already treats it as the judge rather than the generator. And the Fast Models Slow Evidence paper from last week measured that decision-model gates can be 98% accurate on the gate while the end-to-end agent sits at 58%, which is the exact case for making the gate a separate cheap call instead of asking the big model to grade its own work.
What to watch is whether OpenAI publishes calibration numbers. Jev, Strands and Clef all report Brier scores on JevBench. The Decisions docs promise probabilities but, so far, no public measurement of how trustworthy those probabilities are. For a product whose whole value is the number, that is the missing row.
Link: developers.openai.com/api/docs/guides/decisions
← Back to all articles
This is the decision-model category going mainstream at the largest API provider. TypeSafe AI's Jev started the thread. AWS followed with Strands Decider 2B and Cloudflare with Clef, both open weights. Musubi shipped PolicyLM for moderation on Monday. Now OpenAI has the same shape as a hosted product: pick from options, return a calibrated number, do it in one forward pass, make it cheap enough to put in front of every request. The three question types map directly onto what agent harnesses actually need. Predicate for "did this step succeed." Choice for routing. Score for severity or quality gating. Each answer carries a confidence, and refusals come back as a typed answer rather than an error.
The timing with the rest of the week is not an accident. Docker Agent's v1.149.0 release the same day added "OpenAI Decisions as a native evaluator backend," so one harness already treats it as the judge rather than the generator. And the Fast Models Slow Evidence paper from last week measured that decision-model gates can be 98% accurate on the gate while the end-to-end agent sits at 58%, which is the exact case for making the gate a separate cheap call instead of asking the big model to grade its own work.
What to watch is whether OpenAI publishes calibration numbers. Jev, Strands and Clef all report Brier scores on JevBench. The Decisions docs promise probabilities but, so far, no public measurement of how trustworthy those probabilities are. For a product whose whole value is the number, that is the missing row.
Link: developers.openai.com/api/docs/guides/decisions
Comments