Google introduced the Decisions API for Gemini on November 11, 2026. The associated model, Gemini Decision 1, does not write text. It answers predefined questions with yes or no, selects from a list, or assigns a score, and it provides a confidence value for each answer, that is, how reliable the answer is.
The interface is aimed at developers and at anyone who automates workflows with their own scripts. For users of chat and the app, it remains a building block in the background.
What does the Decisions API do differently from a chatbot?
Instead of formulating word by word, the model evaluates the options in a single computation pass. There are three question types: yes or no, selection from 2 to 255 options, and a score on a scale. For each answer, the API returns probabilities and a confidence value between 0 and 1.
According to the documentation, a request may contain up to 1,000 questions about the same input. The input can be text, an image, audio, a video, or a PDF and, together with the questions, may comprise at most 262,144 tokens.
In the blog post, Google puts the response time at 150 to 250 milliseconds. According to its own comparison, the model is almost five times cheaper than Gemini 3.5 Flash-Lite with the Interactions API. The post shows the comparison values only as a chart and cites no independent verification. Google already uses the model itself: in Gemini Notebook, the model routes requests to the appropriate model tier. Google cites everyday requests that are up to 40 percent faster.
What does the Decisions API require, and what does it cost?
You need a Gemini API key from Google AI Studio and a current SDK, that is, a developer package: the documentation specifies version 3.2 or newer for Python and JavaScript. The interface is available only as a beta version (v1beta).
Decisions API: Requirements and limits
As of November 11, 2026, the information comes from Google
- Gemini API key from Google AI Studio
- SDK (developer package) version 3.2 or later, for Python or JavaScript
- Input as text, image, audio, video, or PDF
- No context caching (temporary storage) and no Batch API
- Euro price and availability in the EU not stated
- Gemini in Google Cloud only announced
The interface is available only as a beta version (v1beta).
Sources: Google Blog, Google AI Studio Docs
Google charges only for input tokens and states the price per million in dollars in the blog post. Output tokens cost nothing because the model does not output text. For trying it out, Google AI Studio offers a free quota. Google does not say what usage costs in euros or whether the interface is available in the EU.
Repeated inputs cannot be cached (context caching), and batch requests (Batch API) are not available. As a result, repeated inputs count in full with every call. The API does not accept chat histories; you combine multi-step context yourself as text. If a content filter is triggered, it returns an error code (HTTP 400). According to Google, Gemini in Google Cloud is to follow soon.
What this means for you
Without your own scripts or automations, you can safely skip the Decisions API. If you do write small programs, however, a test in the free quota is worthwhile.
A scenario: for each job you receive several hundred customer photos and want to sort out the blurry ones. You ask the question of whether it is blurry for each photo, and you look at all photos with a confidence value below 0.7 yourself. Google mentions the threshold of 0.7 only as an example. You should calibrate it with your own photos. Whether the model reliably detects blur is something only this test will show.
It is best to bundle your questions into one request, because Google charges for the shared input only once per request. Also plan with dollars until a euro price is available, and check before using customer material whether you are allowed to use the interface in the EU.
Google introduced the Decisions API on November 11, 2026.
Sources
