OpenAI’s Decisions API enters public beta for fixed-answer agent routing
By AgentRiot Editorial
The new endpoint turns text and images into probabilities, categories or rubric scores. Its narrow output contract brings input-only billing to routing work, but evaluation and permissions remain the builder’s responsibility.

An incident report arrives. Before an agent investigates, something has to decide whether it belongs in the outage queue or routine maintenance. Asking a model to write a paragraph is unnecessary if the next step needs only a destination.
OpenAI’s Decisions API now offers a dedicated interface for that kind of work. The company’s October 6, 2026 changelog (opens in new tab) records the beta release; the Decisions guide (opens in new tab) identifies it as public beta, available through POST /v1/decisions with gpt-6-luna, currently the only supported model. General availability is expected “in the coming weeks,” without a firm date.
For agent builders, the useful role is ahead of expensive work: classify an incoming item, then decide which workflow deserves it. Decisions supplies the classification. Your application still controls what happens next.
Three answers, not an open-ended response
A request carries a model, shared input and a list of questions. The guide defines three question types (opens in new tab), each suited to a different downstream need.
A predicate estimates the probability, between zero and one, that a condition is true. An incident system could ask whether a report describes an ongoing service interruption. The answer is an estimate, not an automatic instruction to page an engineer.
A choice selects from supplied options. For that same report, the destinations might be immediate response, scheduled maintenance or manual review. The API reference (opens in new tab) allows string or boolean values and preserves their types: the string "true" and boolean true are distinct. Choice answers include an option-probability distribution and a separate confidence field.
A score evaluates ordered rubric levels. Developers define what makes a report minor or severe, giving adjacent levels distinct observable criteria. Its result is the probability-weighted average of zero-based level indices, so it can fall between levels. That score is not itself a probability; reading it as a percentage would change its meaning.
Independent questions can evaluate the same input in one request. If a later question depends on an earlier answer, OpenAI directs developers to make separate requests.
The narrower interface also draws a useful boundary. Use Structured Outputs with Responses (opens in new tab) for extracted fields or an explanation in a custom JSON schema. Use function calling (opens in new tab) when the model needs to propose a tool and its arguments. A routing label does neither.
Migration means rebuilding the evidence payload
Decisions is not a destination for an unchanged agent conversation. Its input contract (opens in new tab) accepts a text string or user messages containing text and inline images. Non-user roles, tools, files, audio and item references are unsupported. Images must be base64 data URLs, not hosted HTTP links or file IDs, with at most 128 images across a request.
For an incident workflow, assemble the alert text and relevant observations into that permitted input. Keep the evidence sufficient to distinguish the queues. If screenshots matter, fetching and encoding them becomes part of the integration, and part of the time users wait.
Answers arrive in question order. Naming questions makes them easier to associate with application rules. Importantly, a question can return a typed refusal (opens in new tab) while other questions still receive answers. Handle that variant before reading a probability, choice or score. A refusal should go to a fallback path, not be mistaken for evidence that an incident is harmless.
The documented SDK minimums (opens in new tab) include Python 3.26.0 and JavaScript 7.30.0. Check the client version before implementing the new request construction and answer handling.
Input-only billing, with a limited savings claim
The Decisions tariff (opens in new tab) is $0.10 per million input tokens, with no cache-read, cache-write or output-token charges. Regional processing premiums and long-context input multipliers still apply. Other requests using Luna follow their applicable model and processing-tier prices, not this endpoint’s tariff.
Consider a hypothetical million individually short text requests, each billed for 400 input tokens including its question instructions. Decisions costs $40. If the same workload through Responses uses 400 uncached input tokens and 20 billed output tokens per request, Luna’s standard rates (opens in new tab) produce a $50 token bill: $40 for input and $10 for output.
That illustration excludes caching, image inputs, tools, retries, regional and long-context premiums, taxes and integration costs. It assumes identical input usage, rather than proving it. Sending uncertain items to another model adds another bill; measure that fallback traffic before treating the $10 difference as a saving.
OpenAI also claims about 10× faster answers than Responses (opens in new tab). That is a vendor claim, not an AgentRiot performance result or an end-to-end latency guarantee. Image preparation, network time and fallback calls can change the experience.
Test the boundary before letting it act
Start with a routing step whose wrong answers are recoverable. Run the proposed classifier alongside the existing workflow without changing live actions. Compare missed urgent incidents, unnecessary escalations and review volume on labeled examples, including ambiguous reports and out-of-scope requests.
OpenAI recommends setting thresholds from application examples (opens in new tab) according to false-positive and false-negative costs. The distinction has a longer history: scikit-learn’s threshold guidance (opens in new tab) separates predicting a probability from deciding what action to take. A default cutoff need not match the cost of a missed outage.
Do not interpret the separate confidence field as demonstrated accuracy. Keep permissions downstream: selecting an incident queue must not grant an agent access to restart a service. Promote the replacement only when its routing errors, fallback bill and measured total latency justify it. Keep the manual-review route available after rollout.

