OpenAI's Decisions API left public beta on October 22 and is now generally available. It is an interface built for real-time decisions: GPT-6 Luna re-tuned into a "decision model" — given text, images or JSON inputs and preset questions, it returns probability-, option- or score-type answers instead of long generated text.
Three key points:
1. Speed: millisecond responses. A single decision call's latency drops from 1–3 seconds to 100–180 ms, roughly 10x faster than doing the same work through GPT-6 Luna's Responses API, per OpenAI. Early customer data is striking: Datadog reported 40x latency reduction, Box reported 5x speed gains, and language-learning app Speak cut the gap between user actions from 1.4 seconds to 0.13. Product lead Nikunj Handa says the next target is 60 ms — "enough for an API call for every single frame in a video."
2. Cheap: pay only for input. Input costs $0.10 per million tokens, cached input $0.01 per million, and output tokens are free. For high-frequency, low-value tasks like classification, request routing and task prioritization, developers can swap out full LLM calls to save both money and latency.
3. The rival: TypeSafe's Jev. The "decision model" category was created by TypeSafe — founded by a former OpenAI researcher — whose Jev grew into a $7.6B company within weeks. On public benchmarks Jev ranks third while OpenAI's decision model ranks eighth/fourteenth: a classic distribution-vs-quality play, with OpenAI chasing through distribution power.
The GA release adds native GPT-Live voice integration, prompt caching, native JSON inputs, per-question images, up to 400 questions per request, and geo-targeted inference in the US and EU.
Source: read the original. This is a summarized digest for reference only.
Comments