OpenAI Jev: Decisions API, visual inputs, and a refused call

OpenAI Jev is the name this page uses for a search that lands on OpenAI’s Decisions API. On October 6, 2026 that API was in public beta, and a call from this desk returned HTTP 200. The product is powered by GPT-6 Luna. ChatGPT Jev shows up the same way. Tibo Sottiaux, whose bio line reads Codex & ChatGPT @OpenAI, posted it on September 29 at 17:27 UTC, about an hour before the OpenAI Developers announcement already written up on the September 29 page. We did not find a ChatGPT setting named Jev.

Sottiaux’s text says: “Decisions API, for lightning fast constrained decision making powered by Luna. Supports visual inputs, and tuned to be able to make decisions in less than a few hundreds of milliseconds end to end.” The OpenAI Developers follow-up, quoted on the earlier page, says to send text or images as context. A search for Jev visual or Jev image lands on that pair: the posts say the call can take an image. Jev 1.13, as described on this desk, takes a text state.

The Decoder, Jonathan Kemper, September 29, is a page we opened. It says developers define questions with a limited set of predefined answers, pass context as text or images, and get an answer back. The uses it names are classifying content, routing requests, and deciding an agent’s next step. It says the API launched as a limited preview, with broad availability expected in the coming days. The chart on that page is titled Decisions API. Under the title: “Helps apps and agents take action 10x faster than GPT-6 Luna.” The Decisions API bar is labeled 150 ms. The GPT-6 Luna API bar is labeled 1.6 s. The axis is task completion time. The caption says OpenAI says the Decisions API is ten times faster than GPT-6 Luna through the regular API. Sottiaux’s “few hundreds of milliseconds” is a looser wording of a speed claim. We did not time either bar.

The GPT-6 Luna model page, fetched October 2, lists input at $0.1 per million tokens and output at $0.5 per million tokens. Input modalities are text and image. Output is text. Image generation and image edit are marked not supported. That table is the chat model. It does not price /v1/decisions. The image support on the page is input. The Decisions posts describe the same direction: an image goes in, and a selection comes back.

On October 2, 2026 we called POST https://api.openai.com/v1/decisions with a bearer key kept in the environment. An empty JSON object came back HTTP 403. A second JSON object, naming gpt-6-luna and a short routing question, came back HTTP 403 with the same sentence: “Decision API is not enabled for this user.” The error type was invalid_request_error. Three OpenAI-Beta header values, decisions=v1, decision=v1, and decisions-2026-09-29, did not change it. OpenAI had not published a request schema that we could find. This page does not print one.

GET https://api.openai.com/v1/models with the same key returned 141 ids. The list included gpt-6-luna, gpt-6-astra, gpt-6-sol, gpt-6.1-sol, gpt-5.6-luna, and gpt-5.6-sol. No id contained “decision”. POST https://api.openai.com/v1/chat/completions with model gpt-6-luna and a one-line greeting returned HTTP 200. The response model was gpt-6-luna. Usage was 8 prompt tokens and 7 completion tokens, and service_tier was default. We did not time it. That call shows the key can reach Luna chat. It is not a Decisions result, and it is not an image call.

On October 4, 2026 the same key sent an empty JSON object to POST https://api.openai.com/v1/decisions again. The response was HTTP 403, with the same sentence, “Decision API is not enabled for this user.” GET https://api.openai.com/v1/models again returned 141 ids, including gpt-6-luna, and none contained “decision”. We did not repeat the chat completion.

We opened the DevDay 2026 recap that day. The Decisions API section still says: “Available in limited preview today with a broad release planned in the coming days.” The API changelog had no entry for it. The newest items were still September 29: computer use on the Agents API, gpt-6.1-sol, and Ultrafast mode for GPT-6 Astra.

The guide at https://developers.openai.com/api/docs/guides/decisions returned HTTP 404 that day. The reference at https://developers.openai.com/api/reference/resources/decisions returned HTTP 404. The openai-python changelog through 3.24.0, released October 2, lists custom voice creation and agent session events. Those notes do not mention a Decisions method.

OpenAI Developers posted on October 6 at 20:47 UTC that the Decisions API was available to all developers in public beta, and that it makes decisions up to 10x faster than GPT-6 Luna through the Responses API. A follow-up in the same minute names three outputs. Predicates estimate the probability that a statement is true. Choices select from predefined options and include confidence scores. Scores evaluate an input against a numeric range. Another follow-up links the Decisions guide. Tibo Sottiaux posted at 20:56 UTC, in a day-two roundup, “Decisions API live for builders.”

We opened that guide the same day. It says the API is in public beta, and that general availability is expected in the coming weeks. gpt-6-luna is the only model. The endpoint is POST /v1/decisions. A predicate returns a probability from 0 to 1. A choice returns one supplied value. A score is the probability-weighted average of level indices. Choice and score answers also return a confidence field. Images have to be inline base64 data URLs. Hosted HTTP or HTTPS image URLs and file_id inputs are not supported. We did not send an image.

The API changelog for October 6 says the Decisions API was released in beta with gpt-6-luna, and that it turns text and images into typed answers 10x faster than the Responses API. The post’s “10x” is against GPT-6 Luna through the Responses API. The changelog’s “10x” is against the Responses API. The reference page returned HTTP 200. Its title is “Create a decision”. The openai-python changelog for 3.26.0, released October 6, lists standalone Decisions support. We did not install 3.26.0.

Opened October 8, 2026, the guide’s first sentence says the API returns typed answers about 10x faster than the Responses API. The examples ask for these SDK releases or later: Python 3.26.0, JavaScript 7.30.0, Go 3.73.0, Ruby 0.101.0, and Java 4.78.0.

We opened the DevDay 2026 recap again the same day. The Decisions API section still says: “Available in limited preview today with a broad release planned in the coming days.”

On October 6, 2026 the same key sent POST https://api.openai.com/v1/decisions again. The JSON named model gpt-6-luna, one sentence about a double charge, and three questions: a choice among billing, technical, shipping, and other; a predicate; and a score with levels Low, Medium, and High. This page does not print that body. No OpenAI-Beta header was sent. The response was HTTP 200. The model field was gpt-6-luna. The choice was billing, with confidence 1.0 and a probability of 1.0 on billing. The other three probabilities were 0.0. The predicate answer was a probability of 0.98 and had no confidence field. The score was 1.15, with confidence 0.78. Its level probabilities were 0.0 for Low, 0.85 for Medium, and 0.15 for High. Usage was 420 input tokens, 0 output tokens, 0 cached tokens, and 0 cache-write tokens. The response header openai-processing-ms was 354. The clock on this desk, from send to the full body, was 1,858 ms.

An empty JSON object to the same URL, the same day, returned HTTP 400. The error type was invalid_request_error. The message was “Missing required parameter: ‘model’.” The code was missing_required_parameter. GET https://api.openai.com/v1/models again returned 141 ids, including gpt-6-luna, and none contained “decision”. We did not repeat the chat completion.

Later the same day we sent three texts to both hosts, once each. Decisions was POST https://api.openai.com/v1/decisions with model gpt-6-luna. Jev was POST https://api.typesafe.ai/v1/systemone with model jev-latest. Every Jev response named jev-1.13.0. No image went out. The Jev responses carried no processing-time header.

The first text was: “Mochi is curled up in a sunny spot, purring softly and kneading a blanket. Her ears are relaxed and her eyes are half closed.” Both hosts got the instruction “Mochi the cat is relaxed and comfortable.” Decisions took it as a predicate. Jev took it as a noul. Decisions returned probability 0.98, 189 input tokens, and 0 output tokens. The header openai-processing-ms was 407. The clock on this desk was 1,641 ms. Jev returned noul 0.98, 308 input tokens, and 21 output tokens. The clock was 694 ms.

The other two texts are on the TypeSafe quickstart, which we opened in Chrome the same day. The curl sample is one Stripe sentence and one noul, “Does this message express urgency?” The request-body sample adds a choice and a score to that sentence. The playground box is the same sentence. The Python sample repeats the three questions, so we did not send either again. On Decisions, each noul was sent as a predicate. The choice values and descriptions were the ones in the quickstart. Each score level used the quickstart’s string as both label and description, because the Decisions guide has those two fields and the quickstart prints one string per level.

The Stripe sentence was: “Hi, I’ve been trying to connect my Stripe account for 3 days and the integration keeps failing. I’m losing sales. Please help ASAP.” Urgency alone came back 0.99 from Decisions and 0.98 from Jev. Decisions used 182 input tokens and 0 output tokens. openai-processing-ms was 773. The desk clock was 1,596 ms. Jev used 302 input tokens and 21 output tokens. The desk clock was 679 ms.

On the three-question call, both hosts chose technical. Decisions confidence was 0.99, with probabilities 0.0 on billing, 0.99 on technical, and 0.01 on sales. Jev confidence was 0.74, with probabilities 0.17 on billing, 0.83 on technical, and 0.0 on sales. Both scored frustration at 1.0 with confidence 1.0, and put all of the weight on “Frustrated but civil.” Decisions put the urgency predicate at 0.98. Jev put the noul at 0.99. Decisions reported 444 input tokens, 0 output tokens, openai-processing-ms 66, and a desk clock of 476 ms. Jev reported 425 input tokens, 73 output tokens, and a desk clock of 649 ms.

The quickstart prints a response of its own: technical at confidence 0.78, with probabilities 0.85 on technical, 0.0 on sales, and 0.15 on billing, frustration 1.0, noul 1.0, 392 input tokens, and 65 output tokens. The live jev-latest call returned confidence 0.74, probabilities 0.83 on technical, 0.17 on billing, and 0.0 on sales, noul 0.99, and 425 and 73 tokens.

The three yes/no pairs were 0.98 and 0.98, 0.99 and 0.98, 0.98 and 0.99. The department label matched. The confidence and the second-place option did not. One call is not a latency ranking. The Decisions processing header read 407 ms, 773 ms, and 66 ms. The desk clocks on those calls were 1,641 ms, 1,596 ms, and 476 ms. Jev’s desk clocks were 694 ms, 679 ms, and 649 ms.

On October 7, 2026 we sent that Stripe sentence again, ten times to each host, after one warmup call to each host. The warmup is left out of the medians. The hosts alternated which one went first. The clock on this desk runs from send to the full body. No image went out. Jev was jev-latest, and each response named jev-1.13.0. Decisions was gpt-6-luna. The forty measured calls returned HTTP 200. Jev’s responses had no processing-time header. Medians were taken on the unrounded clocks and then rounded to the nearest millisecond.

Urgency alone used the instruction “Does this message express urgency?” Jev’s ten clocks ran from 300 ms to 1,103 ms, median 330 ms. Decisions’ ten clocks ran from 243 ms to 438 ms, median 267 ms. openai-processing-ms ran from 53 ms to 90 ms, median 64 ms.

Jev’s noul stayed 0.98, with 302 input tokens and 21 output tokens. Decisions’ probability was 1.0 on all ten, with 182 input tokens and 0 output tokens. The October 6 single call on this sentence had returned 0.99.

The three quickstart questions went out with the same mapping as on October 6. Jev’s ten clocks ran from 258 ms to 1,063 ms, median 338 ms. Decisions’ ten clocks ran from 255 ms to 515 ms, median 269 ms. openai-processing-ms ran from 57 ms to 322 ms, median 68 ms.

Both chose technical on all ten. Jev’s confidence ran from 0.70 to 0.77. Decisions’ confidence stayed 0.99. Frustration stayed 1.0 on both, with confidence 1.0, and all of the weight stayed on “Frustrated but civil.” Jev’s noul stayed 0.99. Decisions’ predicate probability was 1.0 on all ten. The October 6 single call had returned 0.98. Token counts stayed 425 and 73 on Jev, and 444 and 0 on Decisions.

At the published base input rates, $0.10 and $0.042 per million, the six input counts come to $0.0000189, $0.000012936, $0.0000182, $0.000012684, $0.0000444, and $0.00001785, in the order the calls were made. Both price pages leave output uncharged, so the Jev output tokens are not added. The Decisions guide also names regional premiums. This arithmetic leaves them out. Neither body included an invoice line. The token counts themselves differ, so the two hosts were not billing the same request.

everruns pull 3924, merged September 30, prints the same refusal from its own live smoke: “OpenAI Decisions API error (403 invalid_request_error): Decision API is not enabled for this user.” The pull request says the request and response shape is unverified, because OpenAI had not published a reference, and it keeps the guessed types behind a preview flag. We did not copy those types. The October 6 call above is this desk’s, not a rerun of that pull request.

Jev’s published price on this desk is $0.042 per million input tokens, with output free. The Decisions guide prices gpt-6-luna on /v1/decisions at $0.10 per million input tokens. It says there is no cache-read charge, no cache-write charge, and no output-token charge. Regional processing premiums and long-context input multipliers still apply. The Luna chat prices above are a different list. On September 30 a POST to TypeSafe returned HTTP 200. The October 6 Decisions path returned HTTP 200 on a different host. The two products are separate calls.

Kai Wang’s October 1 post says the API returns one pick plus a confidence score, takes images, and answers in about 150 ms. The October 6 choice answer included confidence 1.0. The score answer included confidence 0.78. The predicate answer had a probability and no confidence field. The 354 ms figure is the processing header on this call. It is a different number from the 150 ms label on the Decoder chart, and from the 1,858 ms clock on this desk. We still did not send an image.

One more path uses the same letters. Suzuki’s October 1 post describes SGLang’s /v1/decisions with a Qwen model reading a game screen, and says that demo picks the next action in under 100 ms. We did not open a repository and we did not run it. OpenRouter’s Decisions API and Venice’s Decisions API, already on this desk, are Jev doors. The SGLang path is a local server. None of those three is the OpenAI call that returned HTTP 200 on October 6.

On October 7, 2026 we opened the JevBench API board. The headline release is v1.6.1 and the revision is v1.7.12. OpenAI Decisions, gpt-6-luna, is composite rank 5 at 62.5, with a 95% interval of 51.6 to 64.6. The capability list prints 73.5, from Intelligence 56.9 and Calibration 90.1, at $0.052 per 1,000 decisions and p50 0.30 s. The v1.7.10 note says the row answered 598 of 600 items on sealed set A5 plus the 300 public items, and that the score is equated. The list price in that note is $0.10 per million input tokens. Jev 1.13.0 on the same board is composite rank 3 at 71.5 and capability rank 2 at 77.1. We did not rerun v1.6.1.

On October 8, 2026 we sent the Stripe sentence again, ten measured calls after one dropped warmup, in two separate passes. Jev ran first. Decisions ran after that pass had finished. The clock on this desk still runs from send to the full body. Medians were taken on the unrounded clocks and then rounded to the nearest millisecond. No image went out.

The Jev pass started at 2026-10-08T12:12:55Z and finished at 12:13:26Z. The model field was jev-latest. All twenty measured responses named jev-1.13.0 and returned HTTP 200. Urgency alone stayed noul 0.98, with 302 input tokens and 21 output tokens, on all ten. Those clocks ran from 628 ms to 3,071 ms, median 668 ms. There was no processing-time header. The three quickstart questions used 425 input tokens and 73 output tokens on all ten. Those clocks ran from 649 ms to 3,591 ms, median 719 ms. All ten chose technical. Sorted, the ten confidence values were 0.68, 0.70, 0.72, 0.73, 0.73, 0.73, 0.74, 0.74, 0.74, and 0.76. Probabilities ran from 0.79 to 0.84 on technical and from 0.16 to 0.21 on billing, with sales at 0.0. Frustration stayed 1.0, confidence 1.0, with all of the weight on “Frustrated but civil.” The urgency noul was 0.99 on six calls and 1.0 on four.

The Decisions pass finished at 2026-10-08T12:14:16Z. It was POST https://api.openai.com/v1/decisions with model gpt-6-luna. An empty JSON body returned HTTP 400, missing model, and openai-processing-ms was 1. The urgency warmup returned HTTP 200, then all ten measured calls returned HTTP 200. Probability was 1.0 on all ten, with 182 input tokens and 0 output tokens. Those clocks ran from 324 ms to 574 ms, median 387 ms. openai-processing-ms ran from 137 to 211, median 160. The three-question warmup returned HTTP 200, then all ten measured calls returned HTTP 200, with 444 input tokens and 0 output tokens. Those clocks ran from 323 ms to 446 ms, median 389 ms. openai-processing-ms ran from 143 to 228, median 186. All ten chose technical at confidence 0.99, with probabilities 0.0 on billing, 0.99 on technical, and 0.01 on sales. Frustration stayed 1.0 on “Frustrated but civil.” The urgency predicate was 1.0 on all ten. The October 7 medians stay in the paragraphs above.

Opened the same day, the API board is still release v1.6.1, now revision v1.7.21, and it ranks 26 hosted systems. OpenAI Decisions is composite rank 7 at 62.5. The capability card is still 73.5, and the rank line says capability rank 6 and official composite 7. Jev 1.13.0 is composite rank 4 at 71.5 and still capability rank 2 at 77.1. We did not rerun v1.6.1.

The same day, Hamed Nilforoshan posted a HiringCafe thread. The post’s summary line says the OpenAI call is twice as expensive and 5 to 10 percent worse. The figures the thread prints are Spearman correlations. On query-to-job relevance, Jev is 0.74, OpenAI is 0.71, and Gemini 3.1 FL is 0.63. On resume-to-job, Jev is 0.71, OpenAI is 0.67, and Gemini 3.1 FL is 0.61. The thread does not print the pair count, a price, or a latency. It says HiringCafe has 2.5 million users. The comparison quotes the October 6 OpenAI Developers post. The main post has no image. Two follow-ups in the thread are 2107644661448024403 and 2107644837407404340. We did not rerun the ranking.