OpenAI has taken its Decisions API out of public beta. The company announced on October 22 that the endpoint, which answers bounded questions with typed answers instead of paragraphs, is now generally available to every developer with API access, powered by GPT-6 Luna. One request can now carry up to 400 questions, double the beta limit, and OpenAI prices the service at $0.10 per million input tokens with no charge at all for output tokens.
The launch closes a three-week sprint: OpenAI first showed the API in limited preview at its DevDay on September 29, opened a public beta on October 6, and moved to general availability on October 22. The product is built for high-volume, low-latency decision work: customer support routing, content classification, document triage and AI agents picking their next action. It arrives alongside OpenAI’s push to strengthen California’s AI safety law and its $30 billion fundraising push at a $1.25 trillion valuation — the API is one of the revenue levers that valuation is paying for.
What The Decisions API Actually Does
Instead of returning prose, the Decisions API returns a typed answer in one of three forms. A predicate estimates whether a condition is true — is there visible damage in this product photo, for example. A choice selects a category, such as routing a support ticket to billing or shipping. A score applies an ordered level, like grading package damage from none to major. Applications predefine the questions and the candidate answers, and the API fills in the blanks.
The endpoint runs on GPT-6 Luna, the speed-tuned model OpenAI introduced alongside GPT-6 Sol in September, and accepts text, images and structured JSON inputs. Questions that are independent of each other can share one input; questions that depend on earlier answers must go in separate requests, since a single 400-question call evaluates shared evidence rather than running a sequence of dependent tasks.
Speed, Caching And The Price Structure
OpenAI describes the API as up to ten times faster than asking GPT-6 Luna the same questions through the standard Responses API, and product lead Nikunj Handa says latency drops from the one-to-three-second range to 100–180 milliseconds. In demonstrations, individual decisions came back in around 150 milliseconds.
Pricing is built around the input side only. Uncached input costs $0.10 per million tokens, cached input costs $0.01 per million, and output tokens are not charged at all. Prompt caching reuses an unchanged portion of the shared input, such as a reference document or a reference image — but the reusable prefix must contain at least 1,024 tokens, and the question text itself does not count toward that threshold. Cache writes carry separate charges, and regional processing premiums can apply.
The generally available version also adds native JSON input, image instruction support, conversation-history input, usage reporting and Private Link, along with native integration with GPT-Live, OpenAI’s real-time voice system, so applications can invoke decision functions during voice interactions. Regional inference processing is available in the United States and Europe, and eligible customers can use zero data retention.
Early Customers And The Market It Is Chasing
OpenAI named Box, Datadog, Ramp, Sierra and Speak among the companies already using the API, according to Seeking Alpha’s newsroom. The numbers attached to the case studies are aggressive: Speak cut the time between actions from 1.4 seconds to 0.13 seconds, and Sierra now runs under 280 milliseconds at the 95th percentile for roughly 90 percent lower cost per call.
The launch puts OpenAI directly up against TypeSafe’s Jev, the decision model that helped define the bounded-decision category. The Artificial Analysis Decision Index currently ranks Jev at number three and OpenAI’s entry at fourteen, which suggests the speed claim will be tested quickly by independent benchmarks — “ten times faster” is OpenAI’s own benchmark claim against its Responses API, not an independent workload measurement.
Why It Matters
The Decisions API is OpenAI’s bid to own the layer between a model and the software that depends on it. Classification, routing and triage are the highest-volume calls many applications make, and every one of them currently pays for a full generative response when it only needs a probability, a label or a score. By selling bounded answers at a tenth of a cent per thousand input tokens with no output charge, OpenAI is turning its own cost advantage into a wedge against every third-party classifier and rules engine in its customers’ stacks.
For developers, the practical guidance from the release is to start with low-authority work: classification, routing, prioritisation. Keep hard authorisation, spending, deletion and safety gates in deterministic application logic, because a typed answer is still a model output, not a guaranteed-correct business decision. Lower latency can multiply a bad rubric just as quickly as a good one.
Background
OpenAI’s API business has been expanding in two directions at once this autumn. The company is selling its flagship models through Google Cloud’s Model Garden after the end of its Microsoft exclusivity era, it introduced usage-policy controls requiring model routers and AI gateways to get its approval, and it has been pushing agentic products like Codex and ChatGPT Work deeper into the enterprise. The Decisions API fits that pattern: smaller, cheaper, faster endpoints that keep developers inside OpenAI’s infrastructure rather than building their own pipelines on top of raw model access.
Source: Search Engine Watch ; Unite.AI
0 Comments