OpenAI's Decisions API, unveiled on 29 September 2026, is built to do one job: pick an answer from a list you give it, fast, with a confidence score attached, instead of writing you a paragraph. If you've already built or bought a "decision layer" for your chatbot or WhatsApp flow — routing messages before a full model call — this is OpenAI folding that exact pattern into its own stack.
What Is the OpenAI Decisions API?#
At its 29 September 2026 DevDay, OpenAI launched the Decisions API in limited preview, running on GPT-6 Luna, its smallest and cheapest model. It's explicitly not a chat model. You send it a message plus a set of labels you define, and it scores each one instead of writing prose back. A label set for a typical WhatsApp intake flow might look like:
- "pricing question"
- "support ticket"
- "escalate to a human"
- "spam or off-topic"
It returns a confidence score against each label in roughly 150 milliseconds, versus about 1.6 seconds for a normal Luna chat completion, per The New Stack's reporting. Pricing wasn't disclosed at launch. The same keynote introduced GPT-6.1 Sol at a fifth of GPT-6 Astra's price and gave the Agents API computer-use and multi-agent support, per OpenAI's own DevDay recap.
Does This Replace the Decision Layer We Wrote About Three Weeks Ago?#
Not exactly, and that's the interesting part. We covered GPTBots.ai shipping a chatbot decision layer built on Jev, TypeSafe's model, which claimed a 445x cost cut over routing every message through a full LLM call. OpenAI's Decisions API chases the same idea — classify first, generate only when you actually need to — except now it's a first-party endpoint instead of a third-party model you have to wire in yourself.
The difference that matters: a Jev-style setup needs you to define, and often retrain, a classifier. OpenAI's version takes labels straight in the prompt, no retraining step — a real win for teams whose routing rules change often.
One caveat: "coming days" is what OpenAI said about broad rollout, and preview APIs slip. I wouldn't rebuild a live flow around this yet.
Who Should Care, and Who Can Skip This#
Running a single-intent bot — book a demo, answer one FAQ, collect a lead? This changes nothing for you today. There's no decision to route in the first place.
If your WhatsApp or web chatbot handles five or more distinct intents before it ever generates a reply — sales question vs. support ticket vs. "get me a human" vs. spam — this is worth watching closely. That's exactly the shape of problem a decision layer solves, and it's the shape of most flows we build for clients across GCC, European, US, Canadian and Indian SMBs.
It's also the same territory Meta's been pushing into with its own enterprise WhatsApp bot platform — routing and cost control are becoming the actual battleground, not raw model quality.
The Honest Trade-Offs#
The upside: a native, fast, cheap classification step means fewer teams need to bolt on a separate routing model just to control costs. 150 milliseconds is fast enough that a routing check stops being a latency tax on the conversation.
The downside: it's a limited preview with no public pricing, no confirmed rate limits, and no word on whether you can tune it against your own labeled data the way a custom classifier lets you.
It's also a single-vendor dependency — if your routing layer sits inside OpenAI's stack, a hiccup there takes down your bot's decisions, not just its replies. Anthropic made a related bet the same week, folding thousands of third-party connectors into one Claude Marketplace instead — different move, same direction: the platforms want more of your stack living inside theirs.
How We're Handling It#
We rebuilt our audit checklist for client WhatsApp flows the week the Jev story broke, and this doesn't replace that checklist — it adds a line to it. For a mutual-funds advisory client's WhatsApp and IVR build, the routing rules we mapped out — after-hours calls deflect straight to WhatsApp, access splits by staff role before a query even reaches a queue — are exactly the narrow, deterministic decisions this API is built for, and that's the shape of flow we're testing it against first.
For the Voiceflow bots we run day to day, including a menswear brand's product-concierge flow, routing already happens through condition blocks before any generative step fires. That's a hand-built version of what OpenAI just shipped as an endpoint. We're not ripping that logic out for a preview API. We're watching whether it gets cheap and stable enough to replace the parts of it that are currently hand-maintained rules instead of a model call.
What I'd Tell a Client Asking About This#
Don't touch a production flow over this yet. If you're already paying for a separate classifier or decision-layer product, put a 30-minute review on your calendar for whenever OpenAI publishes pricing — that number decides whether switching is worth the migration work. If you're designing a new multi-intent bot this quarter, it's worth a look at design time, before you've built the routing logic the hard way.
If you want a second opinion on whether your chatbot even needs a decision layer in the first place — a lot of "AI chatbots" we get called in to fix don't — send me what you're running and I'll tell you straight, no pitch attached. [cal.com/webepex/growth-review]
