OpenAI Decisions API is the quiet DevDay launch that many agency delivery leads will feel first. Announced on September 29, 2026, it asks Luna to pick from answers you define, then returns a choice with a confidence score in roughly 150 milliseconds.

I'm Nitin Kumar, Lead Project Coordinator at Debugged Pro. I help software and IT agencies in India and the USA turn API launches into priced pilots. This briefing sticks to verified reporting from TechCrunch, The New Stack, and OpenAI launch comments, then turns those facts into client-ready steps.

OpenAI related imagery for OpenAI Decisions API agency briefing — Imagery tied to TechCrunch coverage around OpenAI's Decisions API and agent tooling. Image credit: Minh Connors/Bloomberg via Getty Images / TechCrunch

What OpenAI Announced About OpenAI Decisions API

At DevDay, CEO Sam Altman described the API as a way to give Luna a predefined set of options. Those options can be image categories or different agent behaviors. The model stays focused on that choice instead of writing free-form text.

"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," Altman said, according to TechCrunch's report on OpenAI's Decisions API.

The New Stack's Decisions API coverage adds that the product sits on Luna, OpenAI's smallest and most affordable current model. Developers supply the question, the allowed answers, and the context. The API returns one of those answers with a confidence score.

TechCrunch's broader DevDay write-up also notes the API as a real-time decision layer that uses Luna to answer user-defined questions with predefined answers. That matches the constrained-choice story agencies should brief to clients.

If your client is already asking about always-on agents from the same week, keep that thread separate with our OpenAI Dots always-on AI agents agency guide. Decisions API is a routing and classification building block, not a full agent product.

Why OpenAI Decisions API Matters for Agency Delivery

Most agency stacks still call a chat model to pick a label, then parse JSON and retry when the model invents a fourth option. That pattern burns tokens and adds fragile parsers. A constrained decision model is built for the jobs where the answer set is already known.

The New Stack frames three clear jobs:

Classify content into queues you already defined.
Route requests to the right workflow or specialist agent.
Choose an agent's next action from a short policy list.

Those jobs show up every week in India and USA delivery pods. Support triage, sales lead scoring, refund policy branches, and multi-agent handoffs all need a valid enum, not an essay.

OpenAI says the Decisions API returns results in about 150 milliseconds, versus about 1.6 seconds for GPT-6 Luna through the chat API, per The New Stack. OpenAI's Tibo Thibault also said the API is tuned for end-to-end decisions in less than a few hundred milliseconds. That is launch messaging, not a published SLA. Still, it is fast enough that agencies can plan real-time control paths once preview access opens.

OpenAI DevDay 2026 product art related to API and Codex updates — DevDay 2026 product art from TechCrunch coverage of OpenAI API and Codex updates that included Decisions API. Image credit: OpenAI / TechCrunch

OpenAI Decisions API Latency, Inputs, and Preview Status

Launch coverage is consistent on a few hard facts and thin on packaging details. Agencies should brief both sides honestly.

What is confirmed so far

Model: Luna powers the Decisions API in OpenAI and press descriptions.
Inputs: Text or image context, plus a question and finite answers.
Output: One allowed answer plus a confidence score.
Access: Limited preview as of September 29, 2026, with a broader release planned in the coming days.
Platform fit: Mentioned alongside Agents API updates at DevDay.

What OpenAI has not published yet

Pricing for Decisions API calls.
Hard limits on how many candidate answers one request can include.
Tuning options for custom data, if any.
Latency SLOs by region, input size, or concurrency.

An OpenAI spokesperson told The New Stack the company plans to share more "at broad rollout." Until then, do not put Luna chat pricing into a client quote and call it Decisions API pricing. That would be an unsupported assumption.

OpenAI Decisions API vs TypeSafe Jev for Agency Stacks

TypeSafe AI's Jev shipped earlier and became the reference point for fast decision models. TechCrunch notes that Altman's Decisions API aside looks similar: give the model a set of choices and get probabilities or a pick cheaply and quickly.

TypeSafe CEO Diogo Almeida, a former OpenAI engineer, joked on X about "clone wars," according to TechCrunch. He also framed System One style fast decisions as a future path for software automation. That market signal matters even if feature parity is still unclear.

Press comparisons say Jev is strong on text workloads with calibrated probabilities and public pricing. Decisions API coverage highlights multimodal context and OpenAI platform integration. For a USA client already deep in OpenAI Agents API work, Decisions API may be the cleaner ops story once it opens. For a high-volume text-only India ops team optimizing cost today, Jev remains a live alternative to evaluate.

TechCrunch also covers a practical monitoring angle. After agent mishaps on the open internet, OpenAI has described using a separate model to watch for bad actions at significant compute cost. A QueryStory demo used Jev to check each agentic action against the assigned task, blocking high-confidence bad actions and flagging others. The write-up cites monitoring costs of $2.94 with Jev versus $372 with a frontier LLM for that style of check. Those figures are about Jev, not published Decisions API rates. Still, they show why agencies care about cheap decision layers around agents.

OpenAI DevDay interface art used in TechCrunch API coverage — Additional DevDay interface art from TechCrunch's OpenAI API coverage around the Decisions API announcement. Image credit: OpenAI / TechCrunch

Practical Agency Use Cases in India and the USA

USA agencies will hear product and risk questions first. Can Decisions API sit in front of a refund agent. Can it classify screenshots from a computer-use session. Who owns the audit trail when confidence is medium. Bring a one-page decision catalog with allowed answers, owners, and human escalation rules.

India agencies often own the orchestration code, eval sets, and overnight cost dashboards. Build labeled sets for ticket queues, lead routes, and agent next-step policies. Measure wrong-but-valid answers, not only schema failures. A finite answer space removes JSON breakage. It does not remove bad judgment.

Good first pilots look like this:

Support triage into billing, technical, sales, and spam.
Agent handoff among research, ask_user, and stop.
Screenshot routing for computer-use agents that need image context.
Policy gates that only recommend an action while your app enforces permissions.

Keep high-impact actions behind your own authorization layer. A model saying approve_refund is a recommendation. Your service still checks amount limits, account state, and required human approval.

For durable agent infrastructure questions that sit next to this API choice, see our AI agent infrastructure agency guide. Routing speed and workflow durability belong in the same kickoff.

If you need help scoping the pilot, start with our software development consulting page or browse related notes on the Debugged Pro blog.

Agency Checklist: Brief Clients on OpenAI Decisions API

Name the product correctly. OpenAI Decisions API is a constrained decision layer on Luna, not a chat model replacement.
State the access truth. Limited preview now. Broader release promised in the coming days.
Quote latency with sources. About 150ms in launch claims versus roughly 1.6 seconds for Luna chat.
Do not invent pricing. OpenAI has not published Decisions API rates yet.
Define answer sets early. Include an explicit needs_review or unknown path.
Separate recommendation from permission. App policy still owns refunds, sends, and deletes.
Eval on client labels. Track false positives by class cost, not vanity accuracy.
Compare with Jev where needed. Text-only high volume may still favor an existing decision model.
Align India and USA owners. Onshore risk briefing, offshore harness and dashboards.
Wait for docs before hard commits. Preview demos are fine. Fixed-fee cutovers are not.

How Debugged Pro Helps Agencies Act on OpenAI Decisions API

We help agencies turn a preview API into a scoped experiment. That includes decision catalogs, labeled eval sets, latency budgets, and client briefings that sales and engineering can both use.

If a client asks whether to wait for broad rollout, trial Jev in parallel, or wire Decisions API into an existing Agents API path, we start with one reversible job and a clear kill switch. Official claims should stay tied to OpenAI launch materials and high-quality press. Your packaging should stay specific and honest.

Key Takeaways

OpenAI Decisions API launched in limited preview at DevDay on September 29, 2026 as a Luna-powered constrained decision layer.
It returns one developer-defined answer with a confidence score for classification, routing, and agent next-action jobs.
Launch coverage cites about 150ms responses versus roughly 1.6 seconds for Luna through the chat API.
Pricing, option limits, and latency SLOs are not public yet, so agencies should not lock fixed commercial quotes.
India and USA teams should pilot reversible workflows, keep human review paths, and compare OpenAI's option with existing decision models like Jev where cost matters.

FAQ: OpenAI Decisions API for Agency Teams

What is the OpenAI Decisions API?
A limited-preview OpenAI interface from DevDay 2026 that uses Luna to pick from answers you define and return a confidence-scored choice for classification, routing, or agent next steps.

How fast is it?
OpenAI and press coverage cite about 150 milliseconds per decision, versus roughly 1.6 seconds for GPT-6 Luna via chat. Measure your own tails before promising real-time UX.

Is pricing available?
No. OpenAI has not published Decisions API pricing in the launch materials covered by The New Stack and related reporting.

How does it relate to TypeSafe Jev?
Press describes Decisions API as OpenAI's response to the Jev-style decision model category. Jev is already public for text; Decisions API emphasizes OpenAI integration and multimodal context in current coverage.

What should agencies do next?
Brief the preview status, pick one reversible routing or classification job, build labeled evals, keep authorization in the app, and wait for pricing docs before production commitments for India and USA delivery work.

About the Author

Nitin Kumar is Lead Project Coordinator at Debugged Pro. He helps software and IT agencies plan practical AI delivery for clients in India and the USA.