September 30, 2026
AI agents pick their next move faster: OpenAI promises Decisions API responses in 150 ms
OpenAI promises to cut agent action selection time from 1.6 s to 150 ms compared with a standard GPT-6 Luna call.
An AI agent does not always need a detailed response. Decisions API chooses from options defined by the developer to decide where to route a request or what to do next.
Selection instead of generation. In a standard GPT-6 Luna call, the developer waits for the model's response. With Decisions API, they provide text or images, ask questions and define a finite set of answers. Luna chooses from that set: for example, which category a request belongs to.
As of Sep 29, OpenAI had made Decisions API available in a limited preview. The company presented the comparison of 150 ms versus 1.6 s at DevDay, according to The New Stack's report. Here, speed refers to a single decision, not the agent's entire workflow.
Jev from TypeSafe already handles a similar task. The company's founder, Diogo Almeida, introduced it on Sep 15. At OpenAI, he worked on instruction-following methods that became the research foundation for ChatGPT. According to TypeSafe's pricing as of Sep 30, Jev costs $42 per 1 billion input tokens.
Connecting Jev. Get an API key at console.typesafe.ai. Send a POST request to `https://api.typesafe.ai/v1/systemone` with Bearer authentication and JSON containing `state`, `model: "jev-latest"` and `questions`.
For Python, install the SDK with `pip install typesafe-sdk`, set `TYPESAFE_API_KEY` and call `TypeSafeClient().system_one()`. The SDK requires Python 3.10 or later. For coding agents, TypeSafe also offers a skill with integration instructions: `npx skills add typesafe-ai/skills --skill typesafe-ai`. It installs in the project by default; the `-g` flag installs it globally.
OpenAI promised a broad release of Decisions API in the days following the Sep 29 announcement.
