OpenAI launched Decisions API: AI chooses an action instead of a long response

What happened on October 22

OpenAI announced the general availability of Decisions API based on GPT‑6 Luna. The developer sets the question, permissible options, and context; the result can be a choice, an evaluation, or a probability. The company also announced integration with GPT‑Live and caching of repetitive context. These are facts from the official announcement.

OpenAI claims speeds up to ten times faster than GPT‑6 Luna via Responses API. This is a provider comparison for its own terms, not our test and not a prediction of your entire application's acceleration.

Why it's interesting even for a small project

Our conclusion: the architecture itself is what matters here. Decisions can be separated from text generation. For example, a learning service first determines whether a student is asking for a hint or a full explanation, and only then triggers the appropriate scenario. Less work is spent on a response that wasn't needed at all.

Imagine a fictional store: a customer writes "The package arrived, but the lid is missing." An initial category of "packaging issue" is needed, rather than an automatic refund. The system can prepare a request for a photo and pass it to an operator. The payment itself remains a separate action with its own rules.

Three places worth trying

• Support: separating questions about delivery, payment, and defects before preparing a response.
• Educational game: choosing between a hint, repeating a rule, or a new task.
• Work assistant: deciding exactly which data is missing before generating a document.

These are editorial examples of possible applications, not a list of integrations we have launched. For a Ukrainian product, separate examples in Ukrainian are needed: Surzhyk, short messages, errors, and ambiguous phrasing.

How to test without a major rebuild

Take 30 anonymized inquiries and mark the correct route yourself. Add at least five edge cases: two problems in one sentence, irony, empty text, off-topic requests, and messages with conflicting instructions. Record not only the fraction of correct choices but also the consequence of each error.

Compare the full latency from button click to the required action, not just the model's response. For expensive or irreversible operations, provide separate confirmation. Having a confidence score does not, in itself, make an action safe.

What to watch for next

For us, the main question is whether a separate selection mechanism will provide tangible benefit in real Ukrainian requests. A good demonstration does not yet answer this. If you are creating your own agent, start with a single branch and a clear set of errors.

Next step: how to separate AI decisions from API access. To choose a service, there is a catalog of tools by task.

Sources and method

• OpenAI — Decisions API, 22.10.2026

Facts about the services have been verified against official materials. Examples, prompts, and the workflow were prepared by the editorial team. This is not a comparative test of model accuracy or speed. How we work.