Google introduced the Gemini Decisions API on November 11, 2026. The new interface uses Gemini Decision 1 to evaluate predefined possibilities instead of composing a free-form answer. Applications can receive values and probabilities for routing or scoring, but a numerical result does not guarantee that the selected path is correct for every task.
Inputs and outputs
The announcement describes shared multimodal input containing text, images, audio, video and PDF documents, with up to 1,000 questions. This differs from a conversation in which the model writes an extended explanation. Defining the possible answers and the meaning of a score forms part of the developer's task.
The guide defines three types: a binary check, selection from a set and scoring against ordered levels. Questions share the input but cannot see one another's answers. If another question depends on a result already obtained, the application must manage that dependency separately.
That is an important boundary: sending many questions in one request does not automatically execute a sequential plan.
Confidence has a context
Documentation returns probabilities and a confidence value that code can use to choose automatic processing or additional review. It recommends setting thresholds with labelled examples from the application's own tasks, considering the cost of different mistakes.
Editorially, this means a single threshold is not universal evidence of quality. Misrouting an enquiry and incorrectly permitting an action can have different consequences. The system surrounding the model must know which decisions it may automate and which require another check.
Interpreting the latency claims
Google reports 50–100 milliseconds of model compute and 150–250 milliseconds end to end under its stated test conditions. The more detailed comparison concerns inputs up to 16K tokens and between one and 100 questions. Those conditions should not be confused with the maximum of 1,000 questions or the largest supported input.
These are supplier measurements, not a SIExpo benchmark. A real application adds its own network, data preparation, result processing and possibly another model. Its completion time therefore needs separate observation. Faster scoring can improve one stage without establishing the speed of the entire workflow.
Billing is based on input
The Gemini Decision section of the official pricing page lists $0.10 per million input tokens on the paid tier and zero generated output tokens. Shared input is billed once per request, while each question's instructions and criteria add their own tokens. That table lists neither Batch API support nor context caching.
It also distinguishes use of data for product improvement between free and paid tiers. No output charge consequently does not mean zero overall cost or identical data terms.
Integration is only part of the system
LangChain already documents `GoogleGenAIClassifier` as a separate interface for typed answers and probabilities from the Decisions API. This establishes documented integration, not that the editorial team executed it or that every existing conversational wrapper exposes the same features.
A useful evaluation can begin with one concrete choice, realistic examples and independently specified expected outcomes. Accuracy, elapsed time and cases sent for review can then be compared. This makes bounded answers a testable application component rather than another unexplained confidence indicator.
The surrounding workflow still needs to handle unexpected inputs and cases that do not fit the available options. It is useful to record when the application declines to act, rather than counting only successful automatic branches. That provides a clearer view of the trade-off between coverage and mistakes.
The development points towards narrower AI interfaces for particular operations. Their practical value will depend on how well the questions and checks fit the real task. A constrained output format can make integration clearer; evidence from the intended workload remains necessary to decide whether the resulting choices are useful.
Sources: Google — Gemini Decisions API announcement, Google AI Studio — Decisions API guide, Google AI Studio — pricing, Gemini Decision section, LangChain — Google integrations and decision classifier.
The cover is a conceptual AI illustration, not a documentary photograph.
Editorial verification: 2026-11-12T17:35:22.913550+00:00.
