OpenAI's Decisions API is GA at ten cents per million tokens, and it still ranks behind the startups
OpenAI made its Decisions API generally available on October 22, two weeks after the public beta. It is a decision model built on GPT-6 Luna that returns a predicate, a choice or a score instead of prose, and OpenAI says it is up to 10 times faster than calling Luna through the Responses API. Pricing is $0.10 per million input tokens, $0.01 cached, and nothing for output. One request can now carry up to 400 questions, double the beta, and GA adds GPT-Live integration, prompt caching and US or EU data residency. Link: https://openai.com/index/decisions-api/
The pitch is latency. In the TechCrunch piece, OpenAI's Handa describes taking these calls from one to three seconds down to 100 to 180 milliseconds, and says it is just plain old Luna under the hood. Speak, an early customer, cut the time between agent actions from 1.4 seconds to 0.13. That is the real use: the dozens of small calls inside an agent loop, like continue or stop, retry or ask, is this result good enough, where waiting on a full chat completion each time adds up to minutes. TechCrunch: https://techcrunch.com/2026/10/22/openai-makes-its-jev-competitor-widely-available-as-decision-model-competition-heats-up/
The awkward part is in the same article. On the public indices OpenAI's model sits 8th on the AIM System One Index and 14th on the Artificial Analysis Decision Index. Liquid's d1-ultra and Base Labs' elephant-1 lead, and Jev, the startup model that started this category five weeks ago, is 7th and 3rd and still the most queried decision model on OpenRouter.
That is a strange spot for a frontier lab, and it says something about where agent value is moving. The judgment calls inside a loop are a narrow, cheap, latency-bound job, and small specialists are winning it on quality. OpenAI's answer is price and distribution: no output charge and a default inside the API everyone already uses. If that is enough, the startups get squeezed. If the index gap holds, the decision layer becomes the first piece of the agent stack where the lab is not the default choice.
← Back to all articles
The pitch is latency. In the TechCrunch piece, OpenAI's Handa describes taking these calls from one to three seconds down to 100 to 180 milliseconds, and says it is just plain old Luna under the hood. Speak, an early customer, cut the time between agent actions from 1.4 seconds to 0.13. That is the real use: the dozens of small calls inside an agent loop, like continue or stop, retry or ask, is this result good enough, where waiting on a full chat completion each time adds up to minutes. TechCrunch: https://techcrunch.com/2026/10/22/openai-makes-its-jev-competitor-widely-available-as-decision-model-competition-heats-up/
The awkward part is in the same article. On the public indices OpenAI's model sits 8th on the AIM System One Index and 14th on the Artificial Analysis Decision Index. Liquid's d1-ultra and Base Labs' elephant-1 lead, and Jev, the startup model that started this category five weeks ago, is 7th and 3rd and still the most queried decision model on OpenRouter.
That is a strange spot for a frontier lab, and it says something about where agent value is moving. The judgment calls inside a loop are a narrow, cheap, latency-bound job, and small specialists are winning it on quality. OpenAI's answer is price and distribution: no output charge and a default inside the API everyone already uses. If that is enough, the startups get squeezed. If the index gap holds, the decision layer becomes the first piece of the agent stack where the lab is not the default choice.
Comments