Google introduces Gemini Decisions API for evaluating options without generating text

Google announced the Gemini Decisions API on November 11, 2026. Its Gemini Decision 1 model focuses on evaluating predefined options for software rather than composing long answers. Google names model routing, tool selection and rule-based content evaluation as examples. Performance statements are the manufacturer's measurements; Fibona has not independently benchmarked the service.

Probabilities for software rather than a chat response

The model documentation specifies up to 262,144 tokens of shared input and 1,000 independent questions per request. Supported inputs include text, images, video, audio and PDFs. Three question types cover yes/no judgments, selection from a fixed list and rating against an ordered rubric. Selection and rubric questions allow 2–255 options. The responses contain probabilities and confidence values rather than generated prose.

Questions cannot see one another's answers

The API guide explains that questions independently evaluate the shared input. One answer is not passed to another question; developers combine results in application code. It recommends narrow, single-purpose checks. Yes/no probabilities range from 0.01 to 0.99, so thresholds outside that range may never trigger. Python, JavaScript and REST examples are provided, and the quickstart requires SDK version 3.2 or later.

Shared-input costs and question-specific costs

The official price table lists paid usage at US$0.10 per million input tokens. Shared input is billed once, while each question contributes its own instructions and criteria. No output tokens are generated or billed. A free tier is available; Batch API and context caching are not supported for this model. The table says free-tier content may be used to improve products, whereas paid-tier content is not.

The latency figures have a defined scope

Google reports end-to-end latency of 150–250 milliseconds in its measured scenarios with contexts up to 16,000 tokens. That should not be read as a guaranteed response time for every application or input length.

What this means for application design

For a support-ticket application, the approach offers an alternative to generating a paragraph and then extracting a category from it. A confidence score, however, is not a guarantee that a decision is correct. Handling uncertain cases, deciding which actions need human review and testing against real workloads remain separate application-design responsibilities. This is an editorial inference from the product's operation, not a claim that automated decisions are safe in every setting.

Sources: Google's November 11 announcement and the model, API and pricing documentation linked above.

Dji Banner Inside Article - 728x90