OpenAI's Decisions API is generally available: typed answers at $0.10 per million input tokens

OpenAI's Decisions API is now generally available: it takes text, images or JSON and returns a typed answer, either a probability, a pick from your options or a rubric score, from the GPT-6 Luna model. It costs $0.10 per million uncached input tokens and nothing for output. OpenAI moved it out of beta on October 22, 2026.

The reason to care is not the model name. It is that OpenAI is selling a narrow, fast answer instead of a paragraph, and doing so against startups that built exactly that.

What the API does

According to OpenAI's developer documentation, you send a request to POST /v1/decisions with model set to gpt-6-luna, an input that holds the shared evidence and a list of questions. Each question has a unique name, and the response returns an answers array. There are three question types.

• Predicate. Returns a probability from 0 to 1 that a condition is true.
• Choice. Returns one of the values you supplied, plus a probabilities array and a confidence value, for unordered categories.
• Score. Rates the input against 2 to 10 ordered levels and returns the probability-weighted average of the level indices, so the result can fall between levels.

Typical uses in the docs are classifying content, routing requests and prioritizing work. A request can now carry 1 to 400 questions, up from 200 in the public beta. Inputs can be text, messages, tool calls and outputs, images and JSON; files, audio and file ID inputs are not supported.

What changed since the beta

OpenAI's release notes, as logged by Releasebot, list the post-beta additions: native GPT-Live integration for voice apps, prompt caching, native JSON inputs, images in individual question instructions and the doubled question limit. Unite.ai adds that cached prefixes need at least 1,024 tokens and that US and EU inference residency and Private Link support are available.

Price, in context

The docs list uncached input at $0.10 per million tokens and say output tokens are not charged. Releasebot lists cached input at $0.01 per million tokens. Regional processing premiums and long-context multipliers apply, per the docs.

Because the answer is a short typed value, output is trivial and the bill is almost entirely input. That makes cost scale with how much evidence you send, not how long the answer is. For a sense of how this sits against other frontier prices, see our frontier API price tracker.

What OpenAI and customers say

OpenAI says the endpoint can run up to 10x faster than GPT-6 Luna through the Responses API. Customer results listed by Unite.ai, which describes them as customer-stated, include:

• Box: about 5x faster route selection on 30 certificate-of-insurance cases at the same accuracy as the Responses API.
• Datadog: 40x faster p95 latency than available frontier models for scoring command risk, with comparable accuracy.
• Ramp: cost attribution 4x faster and 30 percent cheaper.
• Sierra: under 280 ms p95 for next-action selection.
• Speak: time to next action cut from 1.4 seconds to 0.13 seconds.

These are small samples from vendors and customers who chose to publish. We have no independent replication.

The competition, and what independent indices say

TechCrunch frames the launch as OpenAI's entry into a market for fast, cheap decision models that TypeSafe created with its Jev model. TechCrunch reports TypeSafe is valued at $7.6 billion and that Jev is a deep learning classifier that picks from a set of options and gives a confidence rating.

According to TechCrunch, OpenAI built its model by reconfiguring Luna for faster inference instead of training a dedicated decision model, and product lead Nikunj Handa said speed was the main gain, from one to three seconds down to 100 to 180 milliseconds. On the independent indices TechCrunch cites, OpenAI's model ranks eighth on one and fourteenth on the other, behind offerings from Liquid and Base Lab and behind TypeSafe's latest Jev model. TypeSafe's CEO urged users to test models on their own workloads. That is good advice for any buyer.

Why a typed answer matters for agents

Agents spend much of their time deciding small things: which tool, which route, whether an action looks risky. Using a large model to produce a paragraph for each decision is slow and expensive. A bounded answer is easier to log, threshold and audit. It also fits the control patterns in our explainer on the agent harness.

The caution is calibration. A probability is only useful if 0.8 means roughly four in five. We found no published calibration study of the Decisions API.

How to test it fairly

Pick one real decision your product makes, such as routing a support ticket or flagging a risky command. Collect a few hundred labeled examples, run the same inputs through the Decisions API and through your current approach, and compare accuracy, p95 latency and cost per thousand decisions. Check calibration by bucketing the returned probabilities and seeing whether the 0.8 bucket is right about 80 percent of the time. Add prompt caching if your shared evidence repeats, since the docs say cache hits are discounted and cache writes are charged. Keep a fallback to a larger model for cases where the confidence value is low.

Our position

If you run a classifier or router in a hot path, the API is worth a test, mainly for latency. Build a labeled set, measure accuracy and calibration, and compare against the cheapest alternative you already have. Do not rely on vendor charts, including the ones here. We will update this piece when independent calibration data appears.