OpenAI Decisions API Reaches General Availability With Prompt Caching and 400-Question Batches
OpenAI opened its GPT-6 Luna Decisions API to all customers on October 22, cutting latencies to 100–180 ms and adding $0.01/M cached input pricing alongside Box, Datadog, Ramp, and Speak.
Just five weeks after TypeSafe AI launched Jev, OpenAI brought its competing Decisions API into general availability on October 22—marking the fastest roadmap turnaround for a major AI API product in recent memory.
Built on GPT-6 Luna with 10x faster structured decisions
Powered by a latency-optimized variant of GPT-6 Luna, OpenAI’s Decisions API accepts text, images, JSON, and tool outputs and returns typed predicate, choice, and score answers up to 10x faster than calling Luna through the standard Responses API. OpenAI API product lead Nikunj Handa credited Jev with inspiring the rapid sprint, noting that the endpoint was not even on OpenAI’s roadmap four weeks before its initial preview.
General availability doubled the per-request question limit from beta to 400 questions, added native GPT-Live integration, introduced US and EU inference residency, and rolled out shared-input prompt caching—cutting input costs from $0.10 per million fresh tokens to $0.01 per million cached tokens with zero output charges.
Early customer latency gains
Early enterprise testers included Box, Datadog, Ramp, and AI language-tutoring platform Speak. Speak’s CTO reported that moving action selection onto the Decisions API reduced time between user actions from 1.4 seconds to 130 milliseconds, matching Handa’s reported 100 to 180 ms typical latency range. Even with OpenAI’s broad enterprise reach, Jev still held a lower fresh-token list price ($0.042 vs $0.10/M) and higher scores on early Artificial Analysis Decision Index snapshots.