OpenAI's Decisions API goes GA with 400 questions a call

OpenAI's fast yes-or-no API is out of beta, and it now takes twice the questions per call.

The Decisions API is generally available as of Oct. 22, running on gpt-6-luna, OpenAI's changelog says. It went to public beta on Oct. 6, and we covered that launch. This is the version OpenAI is willing to call finished.

What changed since the beta

Four things, per the changelog.

Each request now takes up to 400 questions, up from 200 in the beta. Shared input is cached, so repeat calls on the same document get prompt caching discounts. Each question can carry its own images in its instructions. And a new input_json part lets you pass structured data without turning it into a string first.

The API itself is unchanged in shape. You send text, images or JSON and ask typed questions, per OpenAI's guide: a predicate that returns a probability, a choice from values you supply, or a score across 2 to 10 ordered levels. OpenAI says it answers about 10 times faster than its Responses API.

The price, from the same guide:

$0.10 per million uncached input tokens. Output tokens are not charged.

Caching kicks in when the shared input is at least 1,024 tokens. Cache hits get discounted input pricing, and cache writes are charged.

The 400 is the upgrade

Speed was already the pitch. The question count is what changes your bill.

Picture a long support ticket you want checked against a policy. In the beta you could ask 200 questions of it in one call. Now it is 400. With the shared text cached, repeat calls on the same ticket pay the discounted cache rate. Output is free.

Yes, it is a GA badge on a product you could already use. But GA is when teams that waited for a stable endpoint put it in production.

It is not the leader. TypeSafe's Jev, which Cloudflare took aim at this month, ranks 7th and 3rd on the AIM System One Index and the Artificial Analysis Decision Index, while OpenAI's model ranks 8th and 14th, TechCrunch reports. OpenAI says its edge is images, with better calibration on image classification.

Where it fits

OpenAI says response times fell from 1 to 3 seconds to roughly 100 to 180 milliseconds, and that trial customers include Box, Datadog, Ramp and Speak, per TechCrunch. Speak said the gap between actions dropped from 1.4 seconds to 0.13 seconds.

Use it where a full chat model is overkill: routing, moderation, tagging, "is this ticket urgent," an agent's next-step choice. Keep the big model for the parts that need prose. Check its probabilities against your own labels before you trust a threshold.

It runs only on gpt-6-luna today, and the endpoint does not accept file_id inputs, per the guide. Send images inline or by URL.

What to watch

OpenAI's stated target is about 60 milliseconds, TechCrunch reports, which would make frame-by-frame video checks practical. Microsoft's Decision-1 and Jev are chasing the same calls.

Run your current classifier's prompts through it this week. Then compare the invoice.