Markdown edition

# Google launches Gemini Decisions API

> A new endpoint scores bounded choices against shared input.

By BIGCHANGE.AI Editorial

Published: 2026-11-11T22:38:37.991Z
Updated: 2026-11-11T22:38:37.991Z
Canonical: https://bigchange.ai/blog/gemini-decisions-api-structured-choices

![A phone photographs a spill.](https://bigchange.ai/api/media/file/spill-camera-concept.png)
AI-generated conceptual illustration for BIG CHANGE

Google introduced the Gemini Decisions API, powered by Gemini Decision 1, on November 11.

## The big change

- The endpoint scores predefined answers rather than writing a prose response. Developers specify the question and use the result to choose what their software does next.
- A confidence estimate is not permission to automate every case. The developer guide recommends setting thresholds with labeled examples from the application’s own traffic and weighing the costs of different errors.
- Multiple questions do not form a conversation. They cannot see each other’s answers, so dependent decisions need application logic or a later request.

## What goes in and what comes back

[Google’s announcement](https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/) describes one shared input of text, images, audio, video or PDFs, with up to 1,000 questions. Types cover yes/no, supplied choices and an ordered scoring rubric.

Its breakroom example pairs a photo with a note, asking about a spill, cleanup action and severity. Those are separate checks on the same material, not evidence that a hazard was reliably detected or resolved.

An API response is information for an application to use. It does not itself establish that a real-world action happened. BIG CHANGE has not tested this API.

## The options define the decision

The [developer guide](https://aistudio.google.com/docs/decisions-api) warns that choice questions distribute probability among supplied options even when none fits. Add a catch-all or a separate scope check. Otherwise, a support router can select a department for a request outside its intended categories.

The guide defines confidence as the estimated probability that the top answer is right. A threshold can send uncertain cases to a person; it needs validation for the intended use. A false positive means treating a condition as present when it is not; a false negative means missing one that is present. Which mistake costs more depends on the task.

For rubric questions, the guide returns a probability-weighted average of level indices. A result can fall between levels; inspect the per-level probabilities when the distinction matters. This is an averaged rating, not necessarily a single category.

## Access, billing and the next step

Google lists AI Studio access, a no-cost prototyping tier and planned Google Cloud support. Pricing is $0.10 per million input tokens, with no output-token charge. Shared input is counted once; each question adds instruction and criteria tokens.

Costs vary with submitted material and question definitions.

OpenAI’s separate [Decisions API](https://bigchange.ai/blog/openai-decisions-api-ga-contract-pricing-limits) was previously covered by BIG CHANGE. Similar names do not establish interchangeable interfaces or prices. Before replacing one service with another, compare its actual input and output contracts and test the cases that matter to the application.

## Sources

- [Google Decisions announcement](https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/) — Launch, model, examples and pricing.
- [Google Decisions API guide](https://aistudio.google.com/docs/decisions-api) — Question contracts, confidence and limits.
- [OpenAI Decisions API coverage](https://bigchange.ai/blog/openai-decisions-api-ga-contract-pricing-limits) — Separate prior product context.