THE SHORT VERSION
● $0.10 per million input tokens. No output charge. No cache-read or cache-write charge either.
● POST /v1/decisions, public beta since 6 October, gpt-6-luna only.
● It returns typed answers, not text — a probability, a choice with a confidence, or a weighted score.
● OpenAI claims about 10x faster than the Responses API. Nobody has measured it independently.
What it actually does
The Decisions API answers questions instead of writing prose. You send an input and a set of questions, each with a name, a type and instructions. You get back an answers array keyed by question name.
Three question types:
- predicate — a probability between 0 and 1 that a condition holds.
- choice — the chosen option, plus a probability for every option and a confidence.
- score — a probability-weighted average of level indices, which can land between levels.
That last one is the interesting one. A score of 2.4 on a five-point scale is a genuinely different output from a model writing "I'd say about a 2, maybe a 3" and you parsing it.
The pricing is the story
| Per 1M tokens | Decisions API | Responses API (same model) |
|---|
| Input | $0.10 | $0.10 |
| Output | $0.00 | $0.50 |
| Cache read | $0.00 | Charged |
| Cache write | $0.00 | Charged |
For classification work the output side is usually small, so zero output cost sounds like a rounding error. It is not, because it removes the variable. A classification pipeline billed only on input has a cost you can compute exactly from your own data before you run it.
Two caveats most coverage leaves out: regional processing premiums and long-context multipliers still apply. The $0.10 is the base rate, not the ceiling.
The limits worth knowing before you build
- Images must be inline base64. Hosted URLs and file IDs are refused. If your pipeline passes image URLs, that is a rewrite.
- One model.
gpt-6-luna only, with no fallback.
- It is beta. Field names and limits may change, and OpenAI expects general availability "in the coming weeks".
- It is not Structured Outputs. This produces answers, not structured objects. If you want a populated JSON schema, that is still the Responses API.
- Zero Data Retention and HIPAA are available to eligible customers, with US and EU data residency.
Three days later, TypeSafe raised $870 million
On 9 October, TypeSafe AI announced an $870 million Series A led by Andreessen Horowitz at a $7.5 billion valuation, with Sequoia Capital, existing investor DCVC and a group of angels participating. Martin Casado joins the board.
TypeSafe makes Jev, which does what the Decisions API does and did it first: type-safe structured values with calibrated probabilities instead of text, trained by a method the company calls Reinforcement Learning for Calibrated Decisions.
| OpenAI Decisions API | TypeSafe Jev |
|---|
| Input per 1M | $0.10 | $0.042 |
| Output | Free | Free |
| Inputs accepted | Text and images | Text only |
| Latency | "~10x faster" (unmeasured) | 70–500ms end to end |
| Status | Public beta | Early access |
TypeSafe says a third of the Fortune 500 already use Jev, and claims 193.6x faster and 444.6x cheaper than conventional LLMs — on its own workflow evals. The company is admirably direct about what that is worth, writing in its own post: "We can't prove it isn't subsidized."
Neither company has published a head-to-head accuracy benchmark. For a product whose entire value is a calibrated probability, that is the number that matters most and the one nobody has.
Where this fits
This is the second structured-decision launch in a week. Liquid AI released d1-3B on 8 October — open weights, typed answers in a single forward pass, zero output tokens. Three products, one shape, seven days.
The pattern is worth naming. For years the industry billed classification as if it were writing, because the same endpoint did both. A model asked "is this invoice overdue, yes or no" generated a sentence and you paid for the sentence. These products stop doing that.
If a meaningful share of your token spend is a model answering questions you could have asked as a dropdown, this is the week to price that work separately.
What to do
- Running high-volume classification on a chat endpoint? Price it against $0.10 input and no output. For most pipelines that is a large cut.
- Need images in the decision? OpenAI only, and base64 inline.
- Cost-sensitive and text-only? Jev is under half the input price, and now has $870 million behind it.
- Before migrating either way, build a labelled set from your own traffic and measure accuracy. No vendor has published a comparison, so this is a test you have to run yourself.
- Do not use either for structured object generation. That is Structured Outputs, and it is a different product.
Sources