OpenAI launches the Decisions API for fast, low-cost model-driven choices

OpenAI's Decisions API is a narrow endpoint. Rather than generating free text, it returns a model-driven choice, such as which branch to take, which tool to call, or whether an item passes a rule, at lower latency and cost than a full completion. OpenAI frames it as a way to make apps more responsive and agent pipelines more deterministic.

The design reflects how agent systems actually spend tokens. Much of an agent loop is routing, classification and gating rather than prose. Moving those steps to a constrained endpoint reduces cost, shortens tail latency and makes outputs easier to validate. That matters given this week's other OpenAI news about agents taking unexpected actions.

Competition in this niche appeared quickly. Fastino's GLiDE v1.2 claims the top spot on a Hugging Face decision-model leaderboard. Perplexity CEO Aravind Srinivas said Perplexity's Decision Model now runs fully locally on a Mac. That gives developers open-weight, on-device alternatives where data must not leave the machine.

What to watch: published pricing and latency against general GPT-6.1 calls, whether decisions come with calibrated confidence (a top HN thread this week argued that LLM confidence scoring is a bad idea), and how the API slots into Open Responses, the Responses-API-derived spec that just moved to the Agentic AI Foundation.