OpenAI's Decisions API reaches Vercel AI Gateway with GPT-6 Luna Decisions

Vercel AI Gateway now routes to OpenAI's Decisions API through an OpenAI-compatible /v1/decisions endpoint, and Vercel's changelog lists GPT-6 Luna Decisions as the flagship model. The change matters for a specific audience: teams that use small, fast classification steps inside agent pipelines - routing support tickets, triaging alerts, enforcing guardrails, scoring outputs against a rubric - and would rather pay for a typed answer than parse generated prose.

Typed answers instead of generated text

Decision models answer structured questions about one shared input and return probabilities, choices, or scores instead of free text. In one request you can mix predicate (yes/no), choice, and score questions; each question's name maps to a key in the returned answers array. OpenAI's Decisions guide describes the same design and claims evaluation roughly 10x faster than the Responses API, with a Playground for testing questions before writing code. The Decisions API is in public beta, and OpenAI says it expects general availability in the coming weeks.

GPT-6 Luna Decisions is a separate model ID

openai/gpt-6-luna-decisions is distinct from openai/gpt-6-luna, which stays a text-generation language model. Requests sent to the decisions ID go to the Decisions API and are billed at its rates - a detail worth checking before you swap IDs in an existing gateway config. Per OpenAI's guide, gpt-6-luna is currently the only supported model.

The Decoder's report on the launch puts pricing at $0.10 per million input tokens with free output tokens, and says OpenAI supports zero-data-retention and HIPAA-compliant workloads; those numbers come from The Decoder's coverage, not from a pricing page cited here. The Decoder also frames the API as OpenAI's response to the "decision models" mini-trend started by Jev in mid-September.

AI Gateway's own decision-model catalog lists Jev, Liquid d1, and Laya alongside GPT-6 Luna Decisions, so one gateway key can compare providers on the same typed-call shape.

Four ways to call it, with minimum SDK versions

The endpoint is reachable through the OpenAI SDK (decisions.create), the AI SDK (experimental_decide, where the yes/no type is called boolean), the existing HTTP API at /v1/evaluate, or Vercel's AI CLI. Versions are a hard gate: the changelog requires OpenAI SDK 7.30.0+ for JavaScript and 3.26.0+ for Python, and experimental_decide needs ai 7.0.128 or later. OpenAI's guide extends the minimums to Go 3.73.0, Ruby 0.101.0, and Java 4.78.0. Older SDK versions simply won't expose the method.

Why gateway routing matters for decision calls

A decision step runs on every ticket, alert, or message - high call volume at small token counts. That makes latency, cost visibility, and fallbacks the deciding factors. AI Gateway advertises one API key across providers, spend tracking, traces on each request, and automatic fallbacks, and Vercel separately ships confidence-based decision fallbacks that escalate to another model when an answer trips a confidence threshold you set.

The practical migration path for anyone already on AI Gateway is short: bump SDK versions, point at the decisions model ID, and keep a text model behind it as the escape hatch.

The launch also sits next to a broader OpenAI API-tier restructuring reported by The Decoder (paid tiers cut from five to three: Build, Launch, and Grow). That tier change was not confirmed in the sources read for this article, so treat it as a lead to verify in your billing console.

Decision endpoints are how agent stacks stop guessing: cheap, fast, typed signals that decide what the expensive model sees next. OpenAI, Cloudflare, and smaller vendors now all sell that layer, and distribution through gateways like AI Gateway and the AI SDK is what turns a launch into a default building block.

Further reading on LinkLoot: AI agent tools guide and AI workflow automation guide.

Source check

Facts here come from the Vercel changelog entry (primary), OpenAI's Decisions API guide, and The Decoder for pricing and tier context. Availability stage as reported by both vendors: public beta on OpenAI, live on AI Gateway, with general availability expected in the coming weeks.