Every SMB eventually automates the same small decisions: routing an email to the right department, qualifying a lead, scoring a support ticket. On September 29, 2026, at its DevDay event, OpenAI unveiled a Decisions API built specifically for this kind of repetitive choice, promising responses far faster than a standard chat call. The tool is promising, but still rough around the edges: its public documentation, pricing, and real-world performance remain unclear three weeks after launch.
In brief
- OpenAI unveiled its Decisions API at DevDay on September 29, 2026, then widened access in a beta release in early October, according to several specialized outlets.
- The API replies with a structured choice (a probability, a selected option, or a score) instead of free text, running on the GPT-6 Luna model.
- OpenAI has mentioned responses in a few hundred milliseconds, but that figure comes from an internal social media post, not from its official documentation.
- Independent testing by specialized outlet eesel.ai in early October 2026 found strong reliability on simple sorting, but gaps on a more sensitive decision.
- As of late October 2026, neither the exact price nor the full technical documentation for the API has been published: treat it with caution before any production use.
What is OpenAI's Decisions API?
The Decisions API is a new technical endpoint, POST /v1/decisions, running on GPT-6 Luna, OpenAI's fast, low-cost model in the GPT-6 lineup. The idea: instead of returning a written answer, it takes context (text or an image) and one or more questions with a limited set of possible answers, then returns the structured answer directly.
Quick definition: a structured response
A structured response is a result encoded in a fixed format (a number, a label, one option from a list) rather than a written paragraph. It is meant to be read directly by another program, without a human rereading it or parsing extra text.
OpenAI describes three possible response types: a probability between 0 and 1 for a given statement, a choice among several options with a confidence level, or a score on an ordered scale of levels. The use cases OpenAI names are content classification, routing requests to the right team, and choosing an AI agent's next action.
A speed claim that is still hard to verify
September 29, 2026
Announced at DevDay
October 2, 2026
Access still closed
October 6-7, 2026
Wider public beta
Late October 2026
General availability claimed, unconfirmed
OpenAI product lead Thibault Sottiaux stated on the social network X that the API returns a decision in "less than a few hundred milliseconds end to end." That figure does not appear on any official OpenAI page reviewed so far: it is a public statement, not a documented commitment.
Pricing calls for the same caution. OpenAI has not published a rate specific to the Decisions API. Only the rates for the underlying model, GPT-6 Luna, are known:
| GPT-6 Luna tier | Input (per million tokens) | Output (per million tokens) |
|---|---|---|
| Standard | $0.10 | $0.50 |
| Batch / Flex | $0.05 | $0.25 |
| Fast | $0.20 | $1.00 |
Source: public GPT-6 Luna pricing, cross-checked via the OpenRouter and Enterprise DNA platforms; no Decisions API-specific rate was published as of late October 2026.
What early independent tests show
Without broad access, specialized outlet eesel.ai tested the GPT-6 Luna model with forced structured outputs, an approach close to what the Decisions API promises, on a sample of 20 support tickets and 160 calls.
On a simple task, routing a ticket to the right team, both models scored perfectly. On a trickier task, deciding whether a reply can go out automatically without human review, the gap widens: the low-cost model with no reasoning made twice as many errors as the more advanced GPT-6.1 Sol model. Based on eesel.ai's estimates using Luna's public token rates, this kind of decision would cost roughly $0.047 per 1,000 tickets with no reasoning, versus $0.089 with a medium reasoning level.
Key takeaway
A very low cost does not guarantee a reliable decision. The higher the stakes of an automated decision (a public reply, a refund, account access), the more sense it makes to pick a more capable model or keep human oversight, even if that costs a bit more per decision.
Why OpenAI is launching this now
This launch did not come out of nowhere. Startup TypeSafe, with its product Jev, already offered a comparable service: accept text, ask a limited-choice question, and return a probability, starting at $0.042 per million input tokens, with free output and a claimed response time of 70 to 500 milliseconds. Jev opened to all developers on September 27, 2026, two days before OpenAI's announcement.
Decisions API (OpenAI)
Jev (TypeSafe)
For an SMB, this competition is good news in the medium term: more players building this kind of building block should mean lower prices and more options. But it is also a reminder that, even before the Decisions API, there are already reliable ways to get a structured answer, for instance through the structured (JSON) outputs of models already available today.
What this means in practice for an SMB
Spot repetitive decisions
Test on your own data
Keep a human safety net
The most promising use cases for an SMB remain low-risk, high-volume ones: sorting a sales inbox, pre-qualifying inbound leads, giving a first priority score to support tickets. For decisions that directly affect a customer or commit the company, caution is still warranted until the tool has stable documentation.
FAQ
What is OpenAI's Decisions API?
It is a new technical endpoint (POST /v1/decisions), unveiled on September 29, 2026, that returns a structured answer (a probability, a choice, or a score) from a context and a limited-choice question, running on the GPT-6 Luna model.
When will it be available to all businesses?
Several specialized outlets report a public beta opening in early October 2026, and some sources mention general availability on October 22, 2026. Neither date is confirmed on OpenAI's official documentation as reviewed in late October 2026.
How much does the Decisions API cost?
OpenAI had not published a rate specific to this API as of late October 2026. Only the rates for the underlying GPT-6 Luna model are known ($0.10 per million input tokens, $0.50 for output at the standard tier), which allows an estimate but not a guaranteed budget.
Should an SMB adopt this tool right now?
For testing and low-risk decisions, yes, it is worth trying. For production use on sensitive decisions, it is better to wait for stable documentation and published pricing, or to rely on already mature solutions such as structured outputs from existing models.
Automating a decision is only worth it if it stays reliable at scale. To go further, read our article on OpenAI's Agents API or explore all our AI resources for SMB leaders.
