Published
Updated
OpenAI's Decisions API preview says results come back in 150 milliseconds
On October 6, 2026 OpenAI Developers posted the Decisions API in public beta, and this desk's POST to /v1/decisions returned HTTP 200. The September 29 post had called it a limited preview for selected API customers. The guide prices gpt-6-luna input at $0.10 per million tokens, with no output charge. The DevDay recap, opened the same day, still said limited preview. TechCrunch on September 30 quotes Sam Altman on speed.
OpenAI Developers posted on September 29 that a Decisions API was in limited preview, powered by GPT-6 Luna. The text says an app defines questions and possible answers, to classify content, route requests, or choose an agent’s next action. The attachment is a video. A follow-up in the same minute says the call can send text or images as context. The example is a support request and the teams it could go to. The API returns a selection the app can use. Preview access is limited to selected API customers for testing. The follow-up says a broad release is planned in the coming days, and it links openai.com/index/devday-2026-recap. We did not open that page.
The New Stack published a write-up the same day. The page fetch failed, so the lines here are the search extract, not the full article. Search metadata dates it 2026-09-29 at 17:15 UTC. The byline in the extract prints “Sep 29th, 2026 1:15pm” and does not name a zone. An OpenAI spokesperson, in that extract, says the company plans to share more “at broad rollout.” The extract repeats the limited preview and the plan for a broad release in the coming days. It also says: “OpenAI says its model returns results in 150 milliseconds, compared to GPT-6 Luna, which would take 1.6 seconds.” Pricing in the extract is still unknown.
The extract’s section heading about confidence scores is the writer’s label. The body line we have does not quote OpenAI on a confidence field, and it does not give a list price for Luna on this endpoint. Veaceslav Rabota’s later post says about 150 ms, roughly 10 times. That sentence matches the New Stack line. It is a recap of the same claim.
Jev’s published price on this desk is $0.042 per million input tokens, with output free. The September 29 extract does not print a price to set beside it. The follow-up says this call accepts images. Jev 1.13, as described on this desk, takes a text state. The Decisions API is a different product, and it is not a route that sends a request to TypeSafe. The call from this desk is on the later page.
TechCrunch, Tim Fernholz, 12:00 PM PDT on September 30, quotes Sam Altman: “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 piece says TypeSafe did not respond to its questions, and that TechCrunch has not spotted developers running the preview. The share image on this page is the still from Diogo Almeida’s post the day before, 18:24 UTC. The post opens with “begun, the clone war has,” then says more competition is good for developers if the model is good, and that building in a system one compatible way is the future. TechCrunch paraphrases that second idea.
The same piece repeats a line Almeida gave TechCrunch the week before, already covered in the September 18 interview: “If you want it really fast and cheap, use dice, right?” The $2.94 against $372 figure in the piece is a monitoring estimate, written up on the Jev Sentinel page. Fernholz’s article does not add a confidence field or a list price. The September 30 reading on this page did not include a call from this desk, and it did not include the recap.
A later page records an October 2, 2026 POST to https://api.openai.com/v1/decisions. The response was HTTP 403, “Decision API is not enabled for this user.” The September 30 reading above did not include that call.
On October 4, 2026 we opened the DevDay recap that follow-up links. The Decisions API section still says available in limited preview today, with a broad release planned in the coming days. The later page records that day’s empty POST to https://api.openai.com/v1/decisions. It returned HTTP 403, “Decision API is not enabled for this user.” The decisions guide and the decisions API reference both returned HTTP 404, and the API changelog had no Decisions entry.
OpenAI Developers posted on October 6 at 20:47 UTC that the Decisions API was available to all developers in public beta. The guide, opened the same day, says public beta, with general availability expected in the coming weeks. It names gpt-6-luna as the only model and prices input at $0.10 per million tokens, with no output charge. The later page records that day’s POST to https://api.openai.com/v1/decisions. It returned HTTP 200, model gpt-6-luna, 420 input tokens, and 0 output tokens. An empty object returned HTTP 400, missing model. The same page later records three texts sent once each to Decisions and to Jev. We opened the DevDay recap again the same day. The Decisions section still says available in limited preview today, with a broad release planned in the coming days.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Primary post is @OpenAIDevs, September 29, 2026, 18:34 UTC. The attachment is a video, so this page has no share image from the post. The text says: give an app real-time decision-making with Decisions API, powered by GPT-6 Luna; define questions and possible answers to classify content, route requests, or choose an agent's next action; available in limited preview.
The follow-up, posted with it, says to send text or images as context. The example is a support request and the teams it could go to. The API returns a selection the app can use. Preview access is limited to selected API customers for testing. Broad release is planned in the coming days. The link is https://openai.com/index/devday-2026-recap/. We did not open that page.
The New Stack URL is https://thenewstack.io/openai-decision-api-luna/. The page fetch failed. Search metadata dates the piece 2026-09-29 at 17:15 UTC. The byline in the extract prints "Sep 29th, 2026 1:15pm" and does not name a zone. The lines we are willing to use from that extract: an OpenAI spokesperson says the company plans to share more at broad rollout; the preview is limited; broad release is planned for the coming days; "OpenAI says its model returns results in 150 milliseconds, compared to GPT-6 Luna, which would take 1.6 seconds"; pricing is still unknown.
The extract's section heading "confidence scores" is the writer's framing. The body line we have does not quote OpenAI saying the API returns a confidence field. The extract does not give a Luna list price for this endpoint. That extract is not a call from this desk.
Veaceslav Rabota's later post says about 150 ms, roughly 10 times. That matches the New Stack sentence. It is a recap, not a second measurement.
TechCrunch, Tim Fernholz, 12:00 PM PDT on September 30, 2026. PDT is UTC-7, so 19:00 UTC. We opened this article. Altman, quoted there: "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 piece says TypeSafe did not respond to its questions. It says TechCrunch has not spotted developers running the preview.
Diogo Almeida, @CompleteSkeptic, September 29, 2026, 18:24 UTC: "begun, the clone war has." The next lines say he loves OpenAI, that more competition and validation is great for developers if the model is good, and that building in a system one compatible way is the future. The attached still is the share image on this page. TechCrunch paraphrases the second idea and prints the same still.
The piece also repeats a line Almeida gave TechCrunch the week before, in the September 18 interview already on this desk: "If you want it really fast and cheap, use dice, right?" The $2.94 against $372 monitoring line in the same piece is on the Jev Sentinel page. We still did not call the Decisions API, and we still did not open the Dev Day recap. The piece does not add a confidence field or a price.
A later page, /news/openai-jev/, records an October 2, 2026 POST to https://api.openai.com/v1/decisions. That call returned HTTP 403, "Decision API is not enabled for this user." The September 30 reading on this page did not include that call.
October 4, 2026. We opened https://openai.com/index/devday-2026-recap/. The Decisions API block still says available in limited preview today, with a broad release planned in the coming days. The later page records that day's empty POST to https://api.openai.com/v1/decisions. It returned HTTP 403, "Decision API is not enabled for this user." The decisions guide and the decisions API reference both returned HTTP 404. The API changelog had no Decisions entry.
October 6, 2026, 20:47 UTC. @OpenAIDevs, status 2107573382229188645, says the Decisions API is now available to all developers in public beta. The guide at https://developers.openai.com/api/docs/guides/decisions, opened the same day, says public beta, general availability expected in the coming weeks, model gpt-6-luna, and $0.10 per million input tokens with no output charge. The later page records that day's POST. It returned HTTP 200, model gpt-6-luna, 420 input tokens, and 0 output tokens. An empty object returned HTTP 400, missing model. The same page later records three texts sent once each to Decisions and to Jev. We opened the DevDay recap again the same day. The Decisions block still says available in limited preview today, with a broad release planned in the coming days.
Compare
Jev's published price on this desk is $0.042 per million input tokens, output free. The September 29 extract does not print a price. The October 6 guide prices gpt-6-luna on /v1/decisions at $0.10 per million input tokens, with no output charge. The September 29 follow-up says the call accepts images. Jev 1.13 on this desk is described as a text state. The two interfaces are different products. Decisions is not a door that forwards a call to TypeSafe.
JevBench's median for Jev 1.13.0 on the v1.5.4 page is 0.62 seconds, a bench measurement of a different task mix. The 150 milliseconds and the 1.6 seconds are the New Stack attribution, not a number we timed.
TechCrunch's September 30 piece quotes Altman on speed and does not print a millisecond figure of its own. Almeida's September 29 post is a reaction, not a timing. The monitoring dollars in that piece are a different story.
Terms
- 150 milliseconds
- The New Stack extract's attribution: OpenAI says the Decisions model returns results in 150 milliseconds, and that GPT-6 Luna would take 1.6 seconds. We did not open the full article. The post itself does not print this pair.
- limited preview
- OpenAI Developers' wording on September 29. Access was limited to selected API customers for testing. The follow-up said a broad release was planned in the coming days. On October 6 the Developers post said public beta for all developers. The DevDay recap, opened that day, still used the September wording.