OpenAI Decisions API: Free Output Tokens, Hosted-Only Weights

OpenAI’s newest API does not generate text at all, and that is the point. The Decisions API returns a probability, a pick from a fixed list, or a score, and OpenAI charges nothing for the output. In under a month, decision models have gone from one startup’s product to a category with at least five serious entrants. OpenAI’s choice to keep its own behind a hosted endpoint may matter more than its price.

What Happened

OpenAI opened the Decisions API to public beta this week, running only on gpt-6-luna. According to The Decoder, it evaluates text, images, or both and returns three output types: yes/no probabilities, picks from predefined categories, and scale-based ratings. Pricing is $0.10 per million input tokens, with output tokens free. The New Stack reports there are no charges for cache reads or writes either.

OpenAI claims the API runs about ten times faster than its Responses API. The company has not published benchmarks or accuracy figures, so that speed claim is unverified. The endpoint supports zero-data retention and HIPAA-compliant use in the US and Europe. Alongside the launch, OpenAI cut its paid API tiers from five to three (Build, Launch, and Grow), with monthly usage limits of $500, $5,000, and $200,000.

Why It Matters

The category started with TypeSafe’s Jev in mid-September. Simon Willison, who had already built a plugin for Jev, described OpenAI’s API as conceptually very similar: both support yes/no, choice, and score questions. The differences are price and inputs. Jev costs 4.2 cents per million input tokens against OpenAI’s 10 cents, but Jev is text-only, while OpenAI’s model accepts images.

The rest of the field went the other direction on deployment. The New Stack lists three open-weight entrants released this month:

Perplexity is the only vendor so far to publish comparative results. It reports an 85.71% overall score across 11 benchmarks against Jev’s 84.51%, though Jev wins six of the eleven individually. Amazon says its 2-billion-parameter model ranks second among public models of that size on JevBench’s public set.

What This Means for Builders

Decision models target the high-volume work that small generative models have handled until now: routing support tickets, classifying documents, flagging damage in photos. Billing only input tokens makes cost depend on how much context a request carries, not on how much the model writes. For classification pipelines, output is a few tokens anyway, so the bigger shift is latency and schema reliability. A decision endpoint cannot return a malformed paragraph when the application expected “billing.”

The same pressure is visible in today’s Claude Haiku 5.5 launch. The New Stack notes Anthropic now pitches its small model at compaction and agentic speed work, partly because decision models are taking over cheap classification.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.