OpenAI Decisions API Goes Generally Available: Input-Only Billing, 150-Millisecond Responses, Redefining AI Decision Costs

On October 22, 2026, OpenAI moved the Decisions API to general availability (GA), ending the public beta that began on October 6. The API does not generate text; it returns only answers, including yes/no, options, or scores. According to the official developer documentation, its response speed is about 10 times that of the standard Responses API, and pricing is $0.10 per million input tokens with no charge for output tokens.

No Generation, Only Judgment: The Price of an Architectural Choice

The core mechanism of the Decisions API is to bypass the text generation stage. A traditional language model API call involves receiving input, activating model reasoning, generating a token sequence, and returning text. The Decisions API skips the final step, and its answers are probability values, enumerated options, or weighted scores.

This choice directly affects latency. Measured data from several early users shows that Box reported about a 5x improvement in routing-selection speed; Datadog reduced p95 latency by a factor of 40; Sierra cut p95 latency to under 280 milliseconds; and Speak reduced next-action time from 1.4 seconds to 0.13 seconds, a 90% reduction.

OpenAI did not train a separate lightweight judgment model for the API. The underlying model remains GPT-6 Luna, and the speed gain comes from skipping the output generation stage. This engineering choice gives the Decisions API the same accuracy ceiling as GPT-6 Luna, and developers do not pay for the compute used to generate text.

Pricing Structure: Why “Input-Only” Is the Key

Standard pricing for GPT-6 Luna is $0.10 per million input tokens and $0.50 per million output tokens. In classification or routing scenarios, the input is usually much longer than the output. Traditional APIs still bill for output tokens. The Decisions API structures outputs as numeric values, which fall outside the billing scope. For classification workloads with very short answers, the actual cost is about one-fifth that of a standard Luna call.

The rate for cached input tokens drops to $0.01 per million, one-tenth the normal input price. A single request supports up to 400 questions, double the public-beta limit of 200.

Three Answer Types: An Underestimated Design Decision

The official documentation defines three question types: predicate returns a probability between 0 and 1, used to determine whether a condition holds; choice returns one item from a predefined enumeration, used for content classification or request routing; and score returns a graded score as a weighted average, used for quality assessment or severity ranking.

Once this three-way classification is standardized at the API level, developers no longer need to use prompt engineering to force the model to answer with just one word and then parse the text. The Decisions API folds the process into the interface design. The three answer types correspond to business logic such as risk-threshold triggers, workflow branching, and priority ranking, matching the internal decision nodes of multi-step AI agents.

Three Weeks from Beta to GA: The Signal Behind the Pace

OpenAI’s release timeline was: showcased at DevDay on September 29, public beta on October 6, and general availability on October 22. During the public beta, it added native GPT-Live integration, Private Link, data residency in the US, Europe, and Asia, organization-level usage reporting, and conversation history support. These features are aimed at enterprise deployments.

The rapid push to GA comes against the backdrop of competitive pressure in the enterprise market. Similar interfaces from AWS Bedrock and Google Vertex AI are competing for users with high-frequency, low-cost AI decision needs. The Decisions API’s HIPAA compliance and Zero Data Retention option target the healthcare and financial industries.

GPT-Live Integration: A Quiet Reshaping of Voice Scenarios

GPT-Live can call Decisions directly or trigger a decision when input or output voice transcripts arrive. This allows systems to route within milliseconds based on real-time transcript segments, affecting the response experience of customer service automation and real-time assistance applications.

Independent Assessment: A Tool, Not a Breakthrough

The Decisions API is an engineering optimization of interface design, not a breakthrough in model capability. The reasoning ceiling of GPT-6 Luna is unchanged. For tasks that require fine-grained semantic understanding or complex reasoning chains, interface-level optimization cannot automatically improve results. Developers still need to design question wording and scoring criteria.

For high-frequency classification, routing, and review scenarios, the API offers a path to lower cost, shorter latency, and more predictable output formats for the same input. Datadog’s 40x p95 latency improvement and Speak’s 90% response-time reduction come from real production systems.

This marks OpenAI’s shift from a model company that provides language generation capabilities to a platform company that provides AI application infrastructure. The combination of the Decisions API, Structured Outputs, and GPT-Live forms the foundational components needed for an agent runtime.