Introduction
Two weeks after Jev introduced the idea of a dedicated "decision model", OpenAI has answered with its own.
At DevDay on September 29, 2026, OpenAI announced the Decisions API: a way to ask a model a question, give it a fixed list of possible answers, and get back a choice, fast.
It is a small announcement with big implications for how AI agents are built.
What Is the OpenAI Decisions API?
The Decisions API is built on a specialized version of GPT-6 Luna.
Instead of writing free text, you define:
- A question
- A bounded set of possible answers
- The context, as text or images
The model returns the chosen answer for each question. OpenAI positions it for three jobs:
- Classifying content
- Routing requests
- Choosing an agent's next action
The Key Claim: Speed
OpenAI says a Decisions call returns in about 150 ms, compared with roughly 1.6 seconds for a standard GPT-6 Luna call. That is about a 10x improvement.
This number comes from OpenAI's DevDay presentation. The conditions behind it were not detailed.
Why speed matters: an agent that makes ten decisions per task spends most of its time waiting on those decisions. Cutting each from seconds to a fraction of a second changes what is practical, from real-time moderation to responsive customer-facing flows.
How It Differs From a Normal API Call
You can already ask an LLM to "answer with one of these options" using structured outputs. The Decisions API aims to make that a first-class, optimized path:
- The answer space is bounded before inference, so the model cannot drift into prose
- It is optimized for latency, not for long generation
- It accepts images, which matters for agents that look at screens
OpenAI is not building a new kind of model here. It is constraining an existing generative model. That is an important distinction when you compare it with purpose-built alternatives like Jev.
What We Still Do Not Know
As of early October 2026, the launch is a limited preview, and much is unpublished:
- Pricing: no Decisions-specific price. The underlying GPT-6 Luna is listed at $0.10 per million input tokens and $0.50 per million output tokens, but the real billing model is unknown
- API contract: no public endpoint, SDK method or response schema
- Confidence scores: it is unclear whether responses include probabilities, and if so how they are produced
- Context limits and benchmarks: not published
- Access: standard API keys are not enabled yet, with a broad release said to be coming
Treat anything beyond the announcement as unconfirmed until OpenAI publishes documentation.
Why It Matters for Marketing Teams
Marketing automation is full of small, repeated decisions:
- Is this comment a complaint, a question or a lead?
- Does this draft respect our brand rules?
- Which of these three hooks should go out first?
- Does this post need human approval?
A fast decision layer lets a system handle these at scale and reserve slower, more expensive reasoning for work that truly needs it.
That is the architecture behind multi-agent AI systems: small specialized steps, each using the right tool.
The Bigger Picture
When both a startup and OpenAI ship the same idea within two weeks, it is a signal.
The industry is converging on a split:
- Generative models to write, design and plan
- Decision models to classify, route and gate
Expect Google and Anthropic to respond. The winning marketing platforms will be the ones that can use any of these models and ground them in a brand's own context.
Related Reading
- Jev vs OpenAI Decisions API: Speed, Cost and Calibration Compared
- What Is Jev? The First System 1 AI Model
- Autonomous Marketing Explained
Frequently Asked Questions
What is the OpenAI Decisions API?
A GPT-6 Luna-based API that returns a choice from a developer-defined set of answers, for classification, routing and agent action selection.
How fast is it?
OpenAI claims about 150 ms per decision, roughly 10x faster than a standard GPT-6 Luna call. This is a vendor claim.
How much does it cost?
OpenAI has not published Decisions-specific pricing. The underlying GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens.
Can I use it today?
It is in limited preview, and standard API keys are not enabled yet.
Does it accept images?
Yes. It supports text and image inputs.
Final Thoughts
The Decisions API confirms that fast, bounded decisions are becoming a core layer of AI agents.
The details that matter most, pricing, schema and confidence handling, are still to come. Until then, it is worth understanding the idea, and watching how it compares with purpose-built options.