OpenAI Launches Decisions API Beta: Typed Answers for Code, 10x Faster

Skip the essay, get a verdict

OpenAI has pushed its Decisions API into public beta — a new endpoint with a deliberately narrow pitch: instead of prompting a chat model for paragraphs of text and then regex-parsing out a true/false or a category label, developers get typed answers their code can branch on.

The endpoint takes text, images, or both, then returns structured, typed responses. OpenAI says it runs about 10 times faster than its existing Responses API for this kind of work. Pricing is $0.10 per million input tokens, with output, cache reads, and cache writes unbilled — an unusual pricing shape that rewards high-volume classification workloads.

The specs

- One model behind it: gpt-6-luna, with an undisclosed parameter count but a 1,050,000-token context window - Hosted only: a single POST to /v1/decisions through OpenAI — no open weights, no self-hosting option - Three question types: every request bundles evidence with named questions, each assigned a type that shapes how the answer comes back - General availability expected in the coming weeks

The context window is a headline feature on its own: over a million tokens of evidence per request means entire document sets, long transcripts, or large image batches can be evaluated in one call.

Why developers should care

This is the first time OpenAI has productized the "prompt, then parse" hack. Every team shipping LLM features that feed into if/else logic — moderation gates, routing decisions, triage classification, fraud flags, content categorization — has been duct-taping this with regex and retries. The Decisions API turns that workaround into a first-class product.

Ten-times-faster matters most for latency-sensitive applications: real-time routing, moderation gates, live triage. Independent benchmarks are still needed before taking the speed claim at face value, but the direction is right.

The pricing comparison is the more interesting signal. Smaller specialized vendors are undercutting OpenAI significantly — TypeSafe's Jev, for example, prices at $0.042 per million input tokens with free output, making OpenAI's base rate roughly 2.4x pricier. Cost-conscious teams will weigh specialized vendors against OpenAI's convenience and image-input support.

The bigger constraint is lock-in: no open weights, no self-hosting — just a hosted POST endpoint. Anyone building decision logic into production pipelines should treat this as a tool to test now, not a default to commit to, until reliability under load is scrutinized outside OpenAI's own numbers.

What this means for your business

1. Your classification pipelines just got simpler. If your team parses LLM output with regex, the Decisions API (or a competitor) eliminates a whole layer of fragile code. Audit where you are doing "prompt, then parse" — that is now a solved problem.

2. Benchmark before you bet. OpenAI's 10x speed claim is self-reported. Run your own latency and accuracy tests against your actual workload before routing production decisions through it.

3. Price the workload, not the hype. At $0.10/M input tokens with free output, high-volume classification is cheap — but specialized alternatives run at less than half the price. For cost-sensitive, text-only workloads, shop around.

4. Do not hard-code a hosted-only dependency without an exit plan. A million-token context window and typed outputs are compelling, but no open weights means you are renting the capability. Keep your decision logic portable so you can switch vendors.

The signal: structured-output APIs are becoming a product category, not a hack. The vendors converging here — OpenAI, TypeSafe, and others — are competing to own the decision layer of the AI stack. Test early, benchmark honestly, and keep your options open.