# Gemini Decisions API turns model output into bounded choices

Google released a specialized endpoint for yes-or-no checks, fixed choices and ordered rubric scores. It returns probabilities instead of generated prose, but teams still need task-specific calibration, fallback rules and human review.[1][2]

Google launched the Gemini Decisions API and Gemini Decision 1 on November 11, 2026. One request can evaluate up to 1,000 independent questions over shared text, images, audio, video or PDFs. The model returns typed probabilities and bills no output tokens.[1][2]

The API fits bounded work such as model routing, policy checks, tool selection and rubric scoring. It does not generate explanations, call tools or perform multi-step reasoning. Developers should keep the Interactions API or another generative model for open-ended work.[2][3]

Google reports 150 to 250 milliseconds of end-to-end latency and a $0.10 input-token price per million tokens. Those are vendor measurements and list prices. Teams need their own held-out cases before choosing thresholds or allowing automatic action.[1][4]

## What Google released

[Google announced [1]](https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/) a v1beta Decisions API powered by the specialized Gemini Decision 1 model. The endpoint evaluates closed sets of outcomes instead of decoding prose token by token.

A request carries one shared input and between one and 1,000 questions. Each question is evaluated independently, so one question cannot read another answer. The shared input can contain text, images, audio, video or PDF content and can reach 262,144 input tokens.[2][3]

The response uses the question keys supplied by the application and returns typed values, option probabilities and confidence. The model produces no text output, which removes output-token charges and avoids parsing prose into application state.[2][3]

## The three question types

Use noul when the application needs a probability for a condition, such as whether a support message describes damaged goods. Use choice when exactly one fixed route should win. Use score when ordered rubric levels represent increasing quality, severity or risk.[2]

Choice and score accept 2 to 255 criteria. Criteria need concrete boundaries. Labels such as good, acceptable and poor leave the model to invent the distinction, while descriptions tied to observable evidence give a team something it can evaluate.[2]

Confidence is separate from the selected value. Your code can accept a high-confidence route, send an uncertain case to a stronger model or ask a person to review it. The threshold should come from held-out cases and the cost of each type of mistake.[2]

## When to use Decisions instead of generation

Use the Decisions API when the outcome is already known as a finite set: choose a queue, check many policies, rank a candidate against a rubric, decide which tool may help or determine whether an agent should stop.[2]

Use a generative API when the system must write an explanation, extract a custom object, reason through several steps or call a tool with arguments. Gemini Decision 1 does not support text generation, thinking, function calling, search grounding, code execution or computer use.[2][3]

A practical system can combine both. A decision call can route a request or identify uncertain cases, and a generative model can handle the smaller set that needs synthesis or an explanation. Evaluate the complete route because a cheap classifier can still raise total cost when it sends too many cases to an expensive fallback.

## A safe evaluation plan

Treat Google’s benchmark and latency claims as reasons to test, rather than expected production results. Build the gate from examples that represent your users, policy and error costs.[1][2]

- Collect a held-out set with clear labels, difficult boundary cases and examples that should go to human review.
- Write criteria from observable evidence and have domain reviewers label disagreements before calling the API.
- Measure accuracy, false positives, false negatives, calibration and latency by subgroup, modality and input length.
- Choose separate confidence thresholds for automation, fallback and human review. High-impact actions should require a person even when confidence is high.
- Compare the full route with the current baseline, including fallback calls, retries, media-token cost and review time.
- Log the model version, criteria, probabilities, chosen route and final reviewed outcome without storing unnecessary sensitive input.
- Re-run the held-out set when criteria, model versions, input formats or downstream actions change.

## Limits that affect system design

The launch supports the Gemini Developer API v1beta and only Gemini Decision 1. It does not support context caching, batch processing, Flex or Priority inference. Repeating the same shared input in later requests bills the input again.[2][3]

Question instructions accept text only. Media belongs in shared input. YouTube and direct Cloud Storage gs:// links are not accepted, so applications must use inline bytes, a supported file URI or an HTTPS URL.[2]

Questions are intentionally isolated. That helps parallel checks, but a later question cannot depend on an earlier answer. Multi-stage decisions need application code or a separate reasoning step.[2]

Google reports strong benchmark calibration, but calibration can shift with a new domain, language, class balance or criteria. A probability should not be treated as a guarantee.

## Frequently asked questions

- Can it replace a normal Gemini model?[2][3] Only for bounded decisions. It cannot write explanations, summarize content, call tools or conduct multi-step reasoning.
- Does zero output tokens mean the request is free?[1][4] No. Google charges for input tokens. Media, long context and repeated uncached input still affect cost.
- Can one call check many policies?[2] Yes. A request can include up to 1,000 independent questions over one shared input.
- Should an application act whenever confidence is high? No. Teams should set thresholds from task-specific evidence and keep human approval for costly, sensitive or irreversible actions.
- Is this the same as structured output?[2] No. Structured output asks a generative model to create data matching a schema. Decisions scores a fixed set of outcomes and returns probabilities.

## Sources

- [[1] Google: Gemini Decisions API launch](https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/)
- [[2] Google AI Studio: Decisions API guide](https://aistudio.google.com/docs/decisions-api)
- [[3] Google AI Studio: Gemini Decision 1 model card](https://aistudio.google.com/docs/models/gemini-decision-1)
- [[4] Google AI Studio: Gemini API pricing](https://aistudio.google.com/docs/pricing#gemini-decision)
- [[5] Gemini API release notes](https://ai.google.dev/gemini-api/docs/changelog)