OpenAI makes Decisions API generally available with 400-question requests

• OpenAI's Decisions API is generally available for bounded text and image decisions, a release included in thsottiaux's shipping roundup.
• Requests now support 400 questions, twice the beta limit; OpenAIDevs' announcement also lists native JSON inputs, prompt caching and GPT-Live integration.
• OpenAI claims up to 10x faster decisions than running the same GPT-6 Luna model through Responses, a comparison WesRoth highlighted.

A JSON key named image_url doesn't fetch an image, and a role key doesn't create a message. Datadog is already scoring agent-generated shell commands before execution, reporting 40x faster p95 than available frontier models with comparable accuracy.

The October 22 changelog records general availability with gpt-6-luna. It is the only supported model at POST /v1/decisions, according to the developer guide.

Applications define questions and allowed answers, then receive one of three types:

• predicate: A probability from 0 to 1 that a condition is true, such as whether a product has visible damage.
• choice: One supplied option, such as continue, retry or clarify, plus probabilities and confidence.
• score: The probability-weighted average of ordered level indices. The result can fall between rubric levels.

The output contract stops there. The guide distinguishes Decisions from Structured Outputs, which generates objects matching a custom JSON schema, and function calling, which requests tool execution with arguments.

Answers can also be refusals. Applications execute any actions selected by a decision.

A request separates shared input from its questions array. Answers return in question order, and question names are echoed back, according to the API reference.

OpenAI's five headline additions since beta are:

• 400 questions per request, up from 200.
• Native JSON input, passed as an input_json content part containing an object or array. Within messages, this is supported only for the user role.
• Images in individual question instructions, allowing separate candidates to be compared against a shared reference.
• Prompt caching, reusing shared input across requests for discounted reads.
• Native GPT-Live integration, allowing decisions during live conversations.

Per-question images are the standout addition: each candidate gets its own question while the reference remains in shared input. Those images count against the same request-level image limits as images in input.

The guide also supports conversation history, including function calls, custom tool calls and their outputs. Images can arrive as public HTTP(S) URLs or inline base64 data URLs; files, audio, item references and file_id image inputs are unsupported.

GPT-Live can call Decisions directly or trigger evaluations as input or output transcripts arrive. OpenAI's avatar example connects three pieces:

1. A trigger of type decisions inside session.triggers.
2. The event session.input_transcript.delta, which supplies transcript updates.
3. A resulting decision.event, from which the application reads the selected reaction and animates the avatar.

Speak reported that the integration reduced time to the next action in its language-learning tests from 1.4 seconds to 0.13 seconds, a reduction of more than 90%.

Customer results published in OpenAI's announcement use different workloads and baselines:

• Box: Approximately 5x faster route selection than Responses on 30 certificate-of-insurance cases, with the same accuracy. Lower-confidence cases went to reviewers.
• Ramp: 4x faster and 30% cheaper AI-cost attribution than GPT-6 Luna, without an accuracy loss.
• Sierra: Under 280 ms p95 for next-action selection in live customer conversations, approximately 3x faster and 90% cheaper per call than its previous approach, with no drop in customer outcomes.

The published base price is $0.10 per million input tokens, or $0.01 per million cached input tokens. Output tokens carry no charge.

Cache writes are billable. The guide's pricing section also notes regional-processing premiums and long-context input multipliers.

The documented deployment and accounting options include:

• Regional processing: United States and Europe, specifically the EEA plus Switzerland.
• Private Link: Requests over a private connection.
• Data controls: Zero Data Retention and HIPAA support for eligible customers.
• Organization usage reporting: Token usage and request counts.