OpenAI's Decisions API Targets Fast Agent Choices, but Details Remain Unpublished

What the Decisions API does

The Decisions API was announced on September 29, 2026. OpenAI's public description is a single paragraph in its DevDay 2026 recap. It says the API focuses Luna's intelligence on "a specific set of user-defined questions with finite pre-defined answers." Developers supply context as text or images and get back an answer they can use to classify content, route requests, or choose an agent's next action.

Access is restricted for now. The recap says the API is "available in limited preview today with a broad release planned in the coming days." OpenAI's developer account added that "preview access is limited to selected API customers for testing."

The problem it targets is familiar to anyone building agents. A simple yes-or-no check, such as whether a shell command is reversible, normally requires a full model round trip: build a prompt, generate tokens, parse the reply, validate its shape and retry if it drifts. Firecrawl argues that a decision-shaped pipeline cuts this to three steps and largely removes the retry branch.

Speed claim and missing details

OpenAI's DevDay slide, as reproduced by The Decoder, reads "Helps apps and agents take action 10x faster than GPT-6 Luna." It shows 150 ms for the Decisions API against 1.6 s for the GPT-6 Luna API. Firecrawl points out that the baseline is OpenAI's own cheapest model, so the figure says little about how the service compares with other vendors. OpenAI product lead Tibo Sottiaux described the latency more loosely as "less than a few hundreds of milliseconds end to end."

Firecrawl reports that, as of September 30, 2026, the Decisions API appears in none of OpenAI's API guides index, endpoint reference index, documentation export, public OpenAPI spec, or the Python, Node and Go SDKs. The only confirmed detail beyond the announcement is image input. The endpoint path, request and response schemas, whether probabilities or confidence are returned, the context limit and pricing are all listed as not published. Vercel, an OpenAI launch partner, noted that the announcement "does not specify whether Decisions API returns option probabilities, rubric scores, or yes-or-no probabilities."

The approach also drew skepticism. A Hacker News commenter argued that such output is "guaranteed to be formatted in a particular way," not guaranteed to be correct, and that grammars on any LLM can achieve the same. Firecrawl counters that structured outputs constrain generated text after the model has decided what to say, while a decisions endpoint bounds the answer space before inference runs.

How Jev compares

Jev, from TypeSafe AI, launched on September 15, 2026, fourteen days before OpenAI's announcement. Firecrawl says the company was founded by former OpenAI researcher Diogo Almeida. Jev is described as a purpose-built "System One" model that takes state in and returns typed decisions, and it cannot generate text. It supports three question types that can be mixed in one call:

• Choice: returns the selected option, a probability for every option, and a confidence value.
• Noul: returns the probability that a condition holds.
• Score: returns a probability-weighted position on an ordered rubric, plus the full distribution.

According to Firecrawl's comparison table, Jev is generally available with full schemas and Python and JS SDKs. It accepts text only, with a 32K context limit on OpenRouter and 64K per TypeSafe's docs. Pricing is $0.042 per million input tokens and $0 for output. OpenRouter telemetry puts its P50 latency at 0.21 s, P95 at 0.34 s and P99 at 0.75 s as of October 1, 2026. For GPT-6 Luna, the table lists $0.10 per million input tokens and $0.50 per million output tokens; no Decisions API pricing exists yet.

Firecrawl treats OpenAI's 150 ms and Jev's measured 210 ms as the same ballpark until someone benchmarks both, since OpenAI's figure is a vendor claim under unstated conditions.

What it means for developers

For teams building agents, the shift is toward treating small control-flow questions as their own class of call, separate from text generation. Firecrawl advises keeping authority in application code: a decision model can judge meaning, but permissions, resource scope, rate limits and business policy should stay in the developer's own policy layer. In its words, a 0.97 on "this looks safe" does not grant an account a permission it did not already have.

Until OpenAI publishes a contract, developers cannot evaluate the Decisions API on pricing, output format or confidence scores. Firecrawl notes that any blog post showing an OpenAI Decisions request body is therefore made up. Jev is the option that can be tested today.