The Decisions API evaluates typed questions against input data using the gemini-decision-1 model with zero text token output.
Traditional generative models produce extensive text strings that require parsing logic to drive application branches. The new endpoint bypasses natural language generation entirely, delivering calibrated probabilities that allow backend code to route, sort, and branch directly on structured outputs.
Each API call transmits a shared input string alongside a structured map of questions. The system processes every question in a single forward pass, maintaining nearly flat latency even as developers scale the number of parallel queries against the same data payload.
The system supports three distinct evaluation modes to handle binary checks, categorical routing, and graded rubrics. Every response type includes a calibrated confidence score ranging from 0.01 to 0.99.
| Question Type | Operational Goal | Returned Metrics |
|---|---|---|
noul |
Check whether a specific statement holds true for the input. | Probability value (0.01 to 0.99), confidence score. |
choice |
Select a single option from a predefined list of criteria. | Top option ID, full probability map, confidence score. |
score |
Rate the input against an ordered rubric of 2 to 255 levels. | Weighted average score, probability map, legend, confidence. |
Binary queries utilize the noul format to return a precise probability that a statement is correct. Choice questions evaluate fixed options, automatically sorting option keys alphabetically in UTF-8 byte order prior to evaluation.
Score questions compute a probability-weighted average across level indices. This returns fractional numerical ratings that map directly back to descriptive rubric tiers via a returned legend object.
Engineering teams can integrate the new functionality utilizing SDK version 3.2 or later via Python or JavaScript installations. Developers execute requests by targeting the v1beta API version and specifying the gemini-decision-1 model identifier.
Google AI Studio reported that output token usage for these decision tasks is permanently pegged at zero. While input tokens accrue based on the size of the shared text payload and question definitions, downstream systems avoid the computational overhead of parsing conversational filler.
Applications can combine atomic checks using ordinary conditional statements and custom weights within local codebases.
