What can the Decisions API do that a normal LLM call cannot?

It mainly returns defined options, scores or probabilities in a typed form rather than free text, and general availability raised the question limit per request from 200 to 400 while adding image, input_json and shared-input caching.

Do I have to move off a generative model with JSON output?

No. OpenAI's guidance in the roundup is that if an existing LLM with JSON output already works well, general availability alone is not a reason to replace it; compare accuracy, behavior on unclear inputs, latency and cost on the same evaluation data.

What else is shutting down for OpenAI API users?

The old Evals API is scheduled to become read-only on October 31 and stop on November 30, and Reusable Prompts and Agent Builder are also slated to stop on November 30.

Large Language ModelsAI Business & IndustryZenn AI/ML26m agoNov 1, 2026, 19:01 JST

OpenAI's Decisions API hits general availability on October 22

Send on LINE

3 Key Points

What happened
OpenAI's Decisions API went generally available on October 22, running on gpt-6-luna and returning typed choices and scores, with the per-request question limit raised from 200 to 400 and support for images, input_json and shared-input caching. Microsoft added its own decision model and an Azure version, while vLLM v0.32.0 added a dedicated endpoint on October 18.
Why it matters
For routing and classification tasks where the options are fixed in advance, a generally available Decisions API becomes a real production candidate to weigh against structured output from a generative LLM on accuracy, latency and cost — though teams already working fine with JSON output have no reason to switch.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Get these free every morning

Context & Analysis

The roundup frames October as the month when options introduced in September began reaching general availability and their operational constraints came into focus. Decisions API is the clearest case: a September limited preview meant evaluation, while the October 22 GA makes it something teams can put into a production comparison for work like inquiry classification and approval routing. The typed-output design is what distinguishes it from asking a generative model for JSON — that line is drawn explicitly, with advice to compare accuracy, behavior on unclear inputs, latency and cost on identical evaluation data.

The surrounding releases show the same pattern across other layers. Microsoft put Microsoft-Decision-1 into Foundry on October 9 and brought Azure OpenAI's Decisions API to Foundry on October 22, and vLLM v0.32.0 added a /v1/systemone endpoint on October 18 for running supported decision models on self-hosted inference. Cloudflare's Clairic, previewed October 19, extracts typed values such as amounts and dates from inputs, which is treated as a different job from choosing among candidates.

What ties these together is the operational fine print rather than the launches themselves. Pricing is shifting — Gemini 4 Argon's introductory price ends in late November and planned increases for Argon and Gemini 3.8 Flash are listed for December 1 and January 1, 2027 — and the deprecation calendar is crowded, with OpenAI's Evals API, Reusable Prompts and Agent Builder all slated to stop on November 30. The practical message is that production comparisons should account for cache write costs, retention settings such as Google's default-on request storage for new paid Gemini API projects, and shutdown dates, not just headline model capability.

FAQ

What can the Decisions API do that a normal LLM call cannot?: It mainly returns defined options, scores or probabilities in a typed form rather than free text, and general availability raised the question limit per request from 200 to 400 while adding image, input_json and shared-input caching.
Do I have to move off a generative model with JSON output?: No. OpenAI's guidance in the roundup is that if an existing LLM with JSON output already works well, general availability alone is not a reason to replace it; compare accuracy, behavior on unclear inputs, latency and cost on the same evaluation data.
What else is shutting down for OpenAI API users?: The old Evals API is scheduled to become read-only on October 31 and stop on November 30, and Reusable Prompts and Agent Builder are also slated to stop on November 30.

Zenn AI/MLRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.