OpenAI Launches Decisions API in General Availability: 10x Faster Verdicts at Zero Output Cost
OpenAI makes its Decisions API generally available with GPT-6 Luna, offering 10x faster typed evaluations at $0.10 per million input tokens and free output tokens.
OpenAI has officially launched the general availability of its Decisions API, powered by GPT-6 Luna. Delivering typed probability, classification, and score verdicts up to 10x faster than standard generation, the endpoint introduces an input-only billing model of $0.10 per million input tokens with zero output token fees, purpose-built for low-latency agentic orchestration.
Input-Only Economics and Zero Output Token Fees
On October 23-24, 2026, OpenAI graduated the Decisions API from public preview to General Availability (GA), deploying it globally via the POST /v1/decisions endpoint. Rather than generating open-ended prose, the API evaluates text, images, and structured JSON to return strictly bounded typed responses: boolean probabilities, choices from predefined sets, or rubric-based integer scores. Crucially, OpenAI established a groundbreaking cost structure: $0.10 per million input tokens, $0.01 for cached inputs, and completely free output tokens, dramatically reducing recurring costs for automated agent routing.
10x Faster Execution and Expanded Production Capabilities
Powered by the compact and efficient GPT-6 Luna model, the Decisions API executes inferences roughly 10 times faster than the standard Responses API, achieving p95 latencies around 150 milliseconds. Since its initial beta, OpenAI has upgraded the service with native voice delegation inside GPT-Live, automated prompt caching, support for inline multimodal image evaluations, and doubled the request capacity to support up to 400 simultaneous questions per API call.
Real-World Adoption Across Agentic and Security Workflows
Early enterprise adopters demonstrate the practical impact of separating lightweight decisions from generative reasoning. Datadog reported a 40x reduction in p95 latency when scoring the security risk of agent-generated terminal commands prior to execution. Similarly, customer service platform Sierra slashed orchestration decision latency below 280 milliseconds at 90 percent lower cost, while language learning app Speak reduced time-to-action from 1.4 seconds to 0.13 seconds by delegating live voice triage to the Decisions API.