Google has introduced the Gemini Decisions API, a new endpoint built for fast, constrained choices rather than ordinary text generation. Announced on November 11, 2026, the API is powered by Gemini Decision 1 and is aimed at tasks such as policy checks, tool selection, routing, ranking and rubric scoring. It accepts a shared multimodal input and can evaluate up to 1,000 independent questions in a single request.
The distinction is important. Many AI applications currently ask a generative model to write a response, then parse that response to decide what software should do next. Google’s new API is designed for the decision itself: yes-or-no conditions, a choice from a fixed list, or a score against an ordered rubric. It returns probabilities and confidence information rather than prose that must be interpreted by another layer of code.
What Google has released
Google describes Gemini Decision 1 as a model that scores candidate answers in one forward pass instead of generating output token by token. The company says the Decisions API is available through the Gemini API in Google AI Studio. The launch post says a request can include text, images, audio, video or PDFs, with up to 256,000 tokens of input and up to 1,000 questions about that shared input.
Google lists three decision types. “Noul” estimates the probability of a yes-or-no condition, optionally with descriptions for true and false. “Choice” selects among two to 255 candidate options and returns normalized probabilities. “Score” evaluates an ordered rubric starting at zero and returns a continuous weighted score with probabilities for the rubric levels.
- Yes or no: Decide whether a stated condition applies.
- Choice: Rank or select a bounded set of possible actions.
- Score: Grade an input against an ordered scale.
Why closed-set decisions deserve their own API
Generative language models are useful when a product needs explanation, drafting, conversation or synthesis. But many production systems need a shorter answer: Is this document missing a required field? Which available tool fits a customer request? Does an incoming message require a human reviewer? Is a claim photograph consistent with a stated category? Those are decision problems, and forcing text generation into the middle can add latency, cost and ambiguity.
A model that returns “probably choose option B, confidence 0.82” can be easier to connect to application rules than one that writes several sentences about its reasoning. Developers can define thresholds, route uncertain cases to a person, or use a cheaper fallback for routine decisions. The model’s output still needs testing and monitoring, but the software interface aligns more closely with the job.
Google says Gemini Decision 1 handles the shared input once and then evaluates multiple questions against it. That is potentially helpful for applications that need several checks on the same call recording, support ticket, inspection image or PDF. It can avoid repeatedly passing the same context into separate generative prompts.
Speed, pricing and the limits of launch claims
Google reports 150 to 250 milliseconds end-to-end latency, with 50 to 100 milliseconds of model compute, for the workload it describes. It lists pricing at $0.10 per million input tokens and no billed output tokens, and says a no-cost tier is available for prototyping in AI Studio. These are Google’s published specifications, not independently audited benchmarks or a guarantee of performance in every region, account tier or workload.
The company also says the model remains nearly flat in latency from one to 100 questions per request under its testing conditions. That makes the product appealing for agent control planes, where software needs to decide whether to call a tool, select a model tier or flag a risky request before an expensive generation step begins.
In real deployments, the full system latency includes network travel, input preparation, policy checks, downstream tool calls and human review. A decision API can reduce one component of a workflow. It cannot by itself make a complex customer process instant or reliable.
Multimodal input changes the practical use cases
The API supports input across text, images, audio, video and PDFs, according to Google. That means a developer can ask a decision question about more than text alone. A service could pair a call recording with account metadata, a scanned claim form with damage photographs, or a video clip with a written policy. The point is not that the model becomes an unquestionable judge. The point is that the application can ask a bounded question over evidence from several formats.
Google’s launch example combines a photo of a breakroom counter with a text note, then asks whether there is an uncleaned spill, what cleanup action fits and how severe the mess is. Such examples are useful for understanding the API shape, but businesses must validate their own data, environments and thresholds. A model suitable for low-risk routing may not be suitable for decisions that affect safety, employment, credit, insurance coverage or legal rights.
How this could fit into AI agents
Agent systems often have a basic loop: inspect a request, choose a tool, take an action, observe the result and decide what happens next. The most expensive part is not always the final answer. It can be the repeated decisions between steps. Should the agent search a database, retrieve a document, request clarification, call a payment service or hand a case to a human?
Google positions the Decisions API for precisely these bounded branches. Its post cites tool selection, policy checks, model routing and context grading. Google also says its own products are using it for model routing, rapid design iteration, reaction selection and adversarial-content detection. Those are vendor-reported deployment examples. They demonstrate intended uses, not a blanket safety certification for every agent workflow.
A disciplined implementation would set a confidence threshold and define a safe outcome below it. For example, a high-confidence policy check might allow a simple next step, a middling score might trigger a second model or additional evidence, and a low-confidence result might go to an employee. That structure treats probability as a signal, not as permission to eliminate accountability.
Calibration is more important than a confident-sounding answer
Google emphasizes calibration, meaning that a probability score should correspond reasonably well to how often a result is correct. If a system says it is 90% confident on a large group of decisions, a well-calibrated system should be right at roughly that rate under comparable conditions. Calibration matters because products need to know when a confidence value is informative enough to automate a low-risk action.
However, calibration can shift when data, users, languages or adversarial behavior change. A development benchmark does not guarantee production calibration. Teams should test on representative data, monitor mistakes and review confidence distributions after release. They should also avoid using a model score as the sole justification for a high-impact decision.
What developers should check before adopting it
The API’s main appeal is efficiency: one shared multimodal input, many bounded questions and no output-token cost. But the integration question is not only technical. A team needs to define each question tightly enough that a correct answer is meaningful, decide how to handle ties or uncertainty, protect customer data and log enough context to investigate a bad result.
A practical adoption checklist
- Start with a low-risk, measurable classification or routing task.
- Write explicit question criteria and test them against representative examples.
- Set confidence thresholds, safe fallbacks and human-review paths before launch.
- Measure latency, accuracy, calibration, cost and the rate of human escalation.
- Review privacy, retention and access controls for every input modality.
Google’s pricing model may be especially attractive for high-volume tasks that previously used generated text only to extract a choice. Yet an apparent cost saving can disappear if poor routing creates downstream errors. A reliable evaluation compares the whole workflow, including error handling and human correction, rather than comparing token prices alone.
Where the API should not be oversold
Gemini Decision 1 does not replace application logic, business rules or human judgment. A closed-set output can be precise only when the allowed options are well designed. If the real-world situation falls outside the choices, the system may still return the least-wrong option. Developers need an “unknown,” escalation or exception path instead of assuming every input fits a predefined category.
Similarly, a confidence score is not an explanation of cause. It may help software select a path, but it does not automatically provide the audit trail that a regulated organization requires. For high-consequence uses, teams may need documented decision policies, independent validation, human sign-off and an ability to contest or correct an outcome.
The broader significance
The launch shows how AI platforms are becoming more specialized. Instead of offering only a general model that generates text, providers are exposing components for individual parts of an application: retrieval, speech, vision, embeddings, safety checks and now structured decisions. That can make systems faster and simpler, but it also asks developers to make more careful choices about where each component belongs.
For agent builders, the practical opportunity is a better control layer between observation and action. For users, the important safeguard is that companies deploy such tools with clear boundaries. Fast decisions are valuable. Fast, unreviewable mistakes are not. The strongest use of an API like this will be in workflows that can state the question clearly, measure the answer and keep people involved when the model’s confidence is not enough.
Sources and publication context
This article was reported and updated on November 13, 2026, from Google’s release post, the Gemini API documentation and Vercel’s availability announcement. Product specifications, pricing and performance figures are attributed to Google unless otherwise stated.
