<- all tokdocs
2026.09.29@mar_antaya

OpenAI's Decisions API Ships as a Limited Preview, Two Weeks After TypeSafe Shipped the Same Primitive

Watch on TikTok

View on TikTok ->

A 19-second clip filmed from the audience at OpenAI DevDay, showing a stage slide that reads "Decisions API / In preview" and a browser-agent demo running against Google Flights. The technical claims in the audio are accurate, including the model name the auto-transcript renders as "Luna," which is real: the endpoint runs on a version of GPT-6 Luna. What the framing leaves out is the chronology. TypeSafe launched Jev, a model that returns typed decisions with calibrated confidence scores, on September 15, 2026. OpenAI announced the Decisions API on September 29, two weeks later. The video presents the capability as arriving, without a competitor in frame.

What is actually on screen

The stage slide reads "Decisions API" with "In preview" beneath it. A presenter in a teal sweater stands in front of a screen replaying a browser session. The prompt at the top of that replay reads "Find a one-way, nonstop flight from SFO to Seattle on Friday after 6 p.m." Below the browser viewport, a caption line tracks what the model chose at each step. One frame shows "Decision: Enter SEA" while the agent fills the destination field. A later frame shows "Decision: Open time filters" while a Google Flights results table is on screen with a time-range slider panel open.

That demo matters more than the slide. The Decisions API is not being shown as a content classifier in the abstract. It is being shown as the step-selection layer inside a computer-use loop, picking the next action from a closed list at each turn.

The primitive: a closed list instead of free text

The audio describes the design precisely. You hand the model "a predefined set of options to choose from," and it returns one. The speaker names three uses: routing a request, classifying an image, and deciding what an agent should do next.

Reporting from the announcement fills in what the clip cuts off. You pass context as text or as an image alongside a closed list of candidate answers. The endpoint returns which option is most likely together with a confidence score. It never generates free text. That constraint is where the speed comes from, and the clip says so directly: "by focusing the model on that choice, we can make it extremely fast while keeping capabilities like image understanding, broad language support, and safety protections."

The speed figure, checked

The clip says "extremely fast" and gives no number. The published figures are 150 milliseconds per decision against 1.6 seconds for a standard call to GPT-6 Luna. Divide those and you get 10.67, which is where the "about ten times faster" phrasing in the coverage comes from. The multiple is real and it is a latency multiple on one model, not a throughput or cost claim.

No pricing has been published. Neither has an accuracy figure, a cap on how many candidate answers a single decision can carry, or whether fine-tuning is available. A confidence score is only useful if it is calibrated, and nothing published so far says how well calibrated this one is. If you plan to threshold on that score and escalate low-confidence cases to a larger model, you are going to have to measure the calibration yourself on your own data.

The chronology the video does not carry

TypeSafe AI released Jev on September 15, 2026. Jev returns typed Choice, Score, and Bool answers with probabilities from a single forward pass and never generates text, and its pitch centers on calibrated confidence rather than overconfident predictions. That is the same shape as the Decisions API: constrained output space, one forward pass, a probability attached.

OpenAI announced the Decisions API on September 29, 2026 at DevDay in San Francisco. Trade coverage framed it as an answer to Jev rather than as a first mover, and the headline on The New Stack's writeup reads "OpenAI answers TypeSafe's Jev with a Decision API built on Luna." A 19-second audience clip with the caption "what I've been waiting for as a dev" is not going to carry that context, and it should not be read as if it does.

What "in preview" means for anything you were about to build

The slide says "In preview" and the announcement described a limited preview with broad availability expected in the coming days. That is not a stable surface. There is no published pricing, so you cannot model unit economics. There is no published SLA on the 150 ms figure, so you cannot plan a latency budget against it. Treat the number as a demo-stage figure until it is under a documented rate card.

The design is worth taking seriously regardless of which vendor you end up on. Routing and next-step selection inside agent loops are the two places where a full generative call is most obviously the wrong tool, because the output is one item from a list you already wrote down. Constraining the output space is the structural fix, and two labs shipping the same fix two weeks apart is the signal to pay attention to.

Key Takeaways

  • The transcript's "Luna" is correct, not a whisper error. The Decisions API runs on a version of GPT-6 Luna.
  • The endpoint returns one option from a developer-defined closed list plus a confidence score. It does not generate free text.
  • Published latency is 150 ms per decision against 1.6 s for a standard GPT-6 Luna call, which is 10.67x, reported as roughly ten times faster.
  • Context can be text or an image, and the on-stage demo uses it as the step-selection layer inside a browser agent, not just as a classifier.
  • TypeSafe shipped Jev on September 15, 2026. OpenAI announced the Decisions API on September 29, 2026. The video's framing implies origination that the dates do not support.
  • Pricing, accuracy, candidate limits, and fine-tuning availability are all undisclosed as of the announcement.
  • Nothing published states how well calibrated the confidence score is, so measure it on your own data before thresholding on it.
  • "In preview" with broad availability promised "in the coming days" means no rate card and no SLA. Do not commit a latency budget to the 150 ms figure yet.

Resources

Published September 29, 2026. Writeup generated from a favorited TikTok.