OpenAI has launched the Decisions API in public beta, a dedicated endpoint at /v1/decisions that accepts text, images, or both and returns one of three structured answer types: a predicate (probability 0–1 that a condition is true), a choice (one pick from a fixed set with per-option probabilities), or a score (probability-weighted average across ordered levels). The only model available is gpt-6-luna, and the endpoint claims roughly 10× speed improvement over the Responses API. The pricing structure is the headline signal. Input tokens cost $0.10 per million. There are no output-token charges, no cache-read or cache-write fees. This is OpenAI explicitly pricing for classification volume — the kind of workload where millions of items pass through a decision gate daily. By eliminating output-token cost entirely, OpenAI undercuts the economics of routing these tasks through general-purpose chat completions, which charge for every generated token even when you only need a single label. Architecturally, the API collapses the generative loop into a discriminative one. Instead of asking a model to generate text and then parsing that text for a label, the endpoint returns typed, structured answers with calibrated probabilities. This is not a new idea — it's the classification head that every ML engineer has bolted onto a foundation model — but packaging it as a first-class API product with its own pricing is a strategic move. It acknowledges that a massive share of production LLM traffic is classification, not generation. The design choices reveal priorities. Images must be inline base64 — no hosted URLs, no file IDs. Multiple independent questions can share one input payload, but dependent questions require separate requests. These constraints suggest an optimized inference path that avoids the overhead of file fetching and maintains request-level isolation for parallelism. The three question types (predicate, choice, score) cover the vast majority of production classification needs without requiring users to design JSON schemas. The competitive implication is straightforward. Every company running content moderation, ticket routing, image inspection, or lead scoring through GPT-4o or Claude is paying generation prices for classification work. The Decisions API offers a purpose-built alternative at a fraction of the latency and cost. This pressures both OpenAI's own Responses API revenue and competitors who haven't yet unbundled their inference pricing by task type. The lock-in vector is gpt-6-luna exclusivity. There's one model, one endpoint, and no option to bring your own weights. If you build classification pipelines on this API, you're coupled to OpenAI's model release cadence, pricing changes, and availability decisions. The $0.10/M input-only pricing is attractive today, but there are no contractual guarantees it stays there post-GA. OpenAI expects general availability in the coming weeks. SDK support spans Python 3.26.0+, JavaScript 7.30.0+, Go 3.73.0+, Ruby 0.101.0+, and Java 4.78.0+. The endpoint supports Zero Data Retention, HIPAA eligibility, and data residency in the US and Europe. These compliance features signal that OpenAI is targeting enterprise classification workloads — healthcare triage, financial document routing, regulated content moderation — where data handling requirements have kept some organizations from adopting LLM-based classification at scale.
