OpenAI opens the Decisions API beta: probabilities, choices and scores at $0.10 per million input tokens, output free; API usage tiers cut from five to three

Why it matters

Key points

On October 6 US time (early October 7 in Beijing), OpenAI opened the Decisions API to all developers as a public beta. It does not write text. It answers three kinds of questions: a predicate returns the probability that a statement is true; a choice picks one of the options the developer supplies, with a probability for each option and a confidence value; a score rates the input against ordered levels the developer defines, returning the probability-weighted average of those levels. Inputs can be text, images or both, and one request can ask several independent questions.

The only model for now is gpt-6-luna, on a dedicated POST /v1/decisions endpoint. OpenAI says it answers about 10 times faster than calling the same model through the Responses API. Input costs $0.10 per million tokens, with no charge for output tokens or for cache reads and writes; regional processing premiums and long-context multipliers still apply. It supports zero data retention and HIPAA use, with data residency in the US or Europe, and OpenAI expects general availability within weeks.

The same day, OpenAI cut its paid API usage tiers from five to three: Build, Launch and Grow. The top tier, Grow, now takes $500 in total payments instead of $1,000 for the old top tier, and paying organizations move to their new tier automatically.

Verifiable

Decisions API guide: https://developers.openai.com/api/docs/guides/decisions . OpenAI API changelog, two October 6 entries (Decisions API beta; usage tiers reduced from five to three): https://developers.openai.com/api/docs/changelog . OpenAI developer forum announcement: https://community.openai.com/t/decisions-api-is-now-available-in-public-beta/1403877 , and the tier change: https://community.openai.com/t/update-to-openai-api-rate-limits/1403852 . Playground: https://platform.openai.com/decisions .

Our take

Three weeks ago, writing about the Jev wave, we argued that production agents do not need yet another chat model so much as a cheap, calibrated, typed decision layer (https://agipulse.org/e/650 ). OpenAI has now shipped almost exactly that layer as a first-class endpoint. That confirms the demand is real, and it means the category will not stay with startups for long. In the same week Liquid AI released open-weight d1 decision models (the 3B d1-3B and an experimental 600M d1-omni that run locally, https://www.liquid.ai/blog/d1-open ). Decision models are becoming a category of their own, contested at once by closed APIs, open local models and specialist startups.

The pricing matters more than the speed. Charging only for input admits that the cost of a decision call sits almost entirely in the context you send; what comes back is a handful of probabilities. The economical pattern for developers becomes pulling routing, moderation, scoring and hand-off-to-a-human checks out of generation calls and sending them here. An agent often makes dozens of such judgments per task, so cutting the cost and latency of each by an order of magnitude changes the bill for the whole chain. Lower tier thresholds point the same way: OpenAI is competing for high-frequency small calls running in production.

Some caveats. OpenAI publishes no calibration metrics and tells developers to set thresholds with their own labelled data, so whether the probabilities are accurate will have to be shown by independent evaluations and users' own data. The 10x comparison is against the same model through the Responses API, not against a dedicated small classifier. There is only one model, and limits on questions per request, input length and dedicated rate limits are not yet documented. For TypeSafe AI and others that got there first with products like Jev, the pressure is on price and distribution; they will need faster responses, private deployment or stronger calibration evidence. What to watch: whether limits and pricing change at general availability, and how well Luna's probabilities match real accuracy in independent tests.