The Decisions API gives applications scores for predefined outcomes. Its practical value depends on how clearly a team defines those outcomes, measures mistakes and handles cases that do not fit.
Read the dated primary source ↓What Google launched
Google introduced the Gemini Decisions API on November 11, 2026, powered by gemini-decision-1 through the Gemini API in Google AI Studio. It supports yes/no checks, selection from fixed choices and ordered rubric scoring over text, images, audio, video and PDFs.
The announcement lists up to 1,000 independent questions per request and a price of $0.10 per million input tokens, with no output-token charges. Shared input is billed once per request; question instructions and criteria still contribute tokens. Google also offers a no-cost tier for prototyping.
Google reports 150–250 milliseconds end-to-end latency in its performance discussion for contexts up to 16,000 tokens, with nearly flat latency from one to 100 questions. That is a reported measurement range, not evidence that the largest supported request finishes equally quickly.
Public record: Google launch announcement — November 11, 2026
The context budget includes your questions
The API documentation specifies 262,144 tokens across the shared input and all questions. It also says questions are isolated: each evaluates the supplied input and cannot inspect another question’s answer. Dependent decisions therefore need sequencing or combination in application code.
A choice question returns a winning option and probabilities. The guide warns that probability is distributed across the supplied options even when none fits. It recommends a confidence threshold, a separate scope check or an explicitly described catch-all option.
The current v1beta endpoint does not support context caching or the Batch API. Repeated input sent across separate requests counts again in full, according to the documentation. Those constraints matter when estimating a recurring classification job.
Public record: Google Decisions API documentation — checked November 12, 2026
Where a useful first trial starts
Our suggested trial is a support inbox with three defined queues and an “other” route. Assemble examples with known destinations, including messages that contain too little information and messages that belong outside the supported categories. Keep the original message beside every result.
Measure incorrect routes, the share sent for review and total latency. Adjust the threshold on a separate set of examples, then test it on messages that were not used to tune it. A high score deserves scrutiny if similar real cases still reach the wrong queue.
Our reading is that this release makes a familiar engineering choice easier to inspect: define the outcomes first, then evaluate whether the model reliably separates them. A working integration needs an explicit response to uncertainty as well as a fast response to ordinary inputs. These are editorial evaluation suggestions; Promptwire has not independently tested the API.
Follow the evidence.
Primary sources checked November 12, 2026. Availability, pricing and performance are attributed to Google. Promptwire has not independently benchmarked the API. Trial suggestions and interpretation are editorial analysis. Cover: AI-assisted conceptual illustration, not a product screenshot.
- Google launch announcement — November 11, 2026
- Google Decisions API documentation — checked November 12, 2026
Public sources support the factual background. Analysis and illustrative exercises are identified in the text. How we assess AI · Editorial standards
Spot a factual error or an outdated detail?
