Google’s Gemini Decisions API is designed for a specific kind of AI task: choosing between defined options quickly and consistently.
Introduced on November 11, 2026, the API is powered by Gemini Decision 1. Google says it can evaluate up to 1,000 yes/no, multiple-choice and rubric-scoring questions in one pass over shared input that may include text, images, audio, video and PDFs.
This is different from asking a general-purpose chatbot to explain a problem in paragraphs. The API is built to return structured decisions and confidence scores that software can use.
What is the Gemini Decisions API?
The Decisions API is a developer endpoint for bounded decisions. A bounded decision has a defined set of possible answers.
For example:
- Does this invoice contain a suspicious charge? Yes or no.
- Which department should receive this support request? Billing, technical support or sales.
- Does this document meet the required quality rubric? Score it against a defined scale.
Google says a request can contain shared multimodal input and up to 1,000 independent questions. The model evaluates the questions in a single forward pass rather than generating a separate long answer for each one.
How is it different from a normal Gemini response?
A standard generative model is designed to produce language. It can explain, draft, reason and respond in open-ended ways.
A decision model is designed to score candidate answers. It returns probabilities for the available choices, which makes it easier for an application to route work or decide whether a human should review a case.
Google describes three decision types:
Yes or no
The model estimates whether a condition is true. This can support checks such as whether a document contains a required field or whether an image shows a defined condition.
Multiple choice
The model selects from a set of options and returns normalised probabilities. This can help route a support ticket or classify a document.
Rubric score
The model assigns a score across an ordered scale. This can be used for quality grading, content review or prioritisation, provided the rubric is clearly defined and validated.
Multimodal input changes the possibilities
The API can process text, images, audio, video and PDFs in a shared input. That matters because many business decisions depend on more than text.
An insurance workflow might compare a claim description with scanned paperwork and damage photographs. A customer support system might combine a call recording with account context. A quality-control application might inspect a video and answer several checklist questions.
The model’s output can include confidence scores from zero to one. Developers can use those scores to decide when to accept a result, request another check or send the case to a human reviewer.
A confidence score is not a guarantee of correctness. It must be calibrated and tested on the actual application.
Performance and pricing
Google’s announcement describes end-to-end latency of approximately 150 to 250 milliseconds for the stated operating conditions. It lists pricing at $0.10 per million input tokens, with no output-token charge for the Decisions API.
Google also says the shared input is billed once per request, while each question adds its own instruction and criteria tokens. The company reports benchmark results comparing the API with other approaches.
These are vendor-reported specifications and benchmark claims. Real latency and cost will depend on input size, deployment conditions, request patterns and current pricing. Developers should verify the latest documentation before estimating production costs.
Five use cases worth testing
- Customer support routing: Classify incoming requests and send them to the right team.
- Document quality checks: Check whether forms contain required information before processing.
- Safety and policy screening: Identify cases that need moderation or additional review.
- Agent tool selection: Help an application choose which tool or workflow should run next.
- Multimedia inspection: Apply a checklist to images, audio, video or PDFs.
The strongest use cases are those where the answer options are clear and the cost of a slow, free-form response is unnecessary.
What developers should not overlook
A fast decision API can make an application more efficient, but it does not remove the need for good system design.
Create representative test data. Measure false positives and false negatives. Set different thresholds when the cost of each error is different. Route uncertain or high-impact cases to people. Keep an audit trail of the input, model output and final action.
Do not use a model score as the sole basis for consequential decisions about employment, credit, healthcare or access to essential services without appropriate safeguards, validation and human accountability.
Gemini Decisions API versus a general-purpose AI assistant
Think of a general-purpose assistant as a colleague who can discuss a problem and write a report. Think of a decision API as a fast evaluator that scores a defined set of options.
They can work together. A decision model can classify or route a case, while a generative model can explain the result or prepare a response. The system should still provide a human path when the model is uncertain.
Why this launch matters
AI applications need both generation and judgement. The Gemini Decisions API reflects growing interest in specialised models that handle classification, scoring and routing without producing unnecessary text.
For developers, the opportunity is to build faster workflows around structured decisions. The responsibility is to prove that those decisions are reliable enough for the task.