The Pulse
Google releases Gemini Decisions API for multimodal choices
The API evaluates up to 1,000 independent questions against a shared input of text, images, audio, video or PDFs. Google says the new Gemini Decision 1 model returns probabilities and confidence scores instead of generating answer text.
Developers can submit up to 1,000 independent questions about a shared input through Google's new Gemini Decisions API, which returns probabilities and confidence scores rather than generating answer text. The endpoint is powered by Gemini Decision 1, a model released through the Gemini API in Google AI Studio, Google said in a November 11 post.
The input can combine text, images, audio, video and PDFs. The service is designed for bounded decisions such as choosing tools, checking policies, routing requests to models and grading context.
Questions and confidence
A request can mix yes-or-no questions, multiple-choice selections and ordered rubric scores. For a yes-or-no question, the API estimates the probability that the condition holds; developers can also provide descriptions of what counts as true or false.
A multiple-choice question accepts 2 to 255 options and returns the selected option alongside probabilities for every candidate. A rubric question produces a weighted score and probabilities for its levels; the documentation specifies 2 to 255 non-empty levels.
Each answer includes a confidence score from 0 to 1. Google's documentation describes it as an estimate of the probability that the top answer is correct and shows how an application could send a lower-confidence answer for human review.
Questions are assessed independently against the shared input: they cannot see one another, and adding or removing a question does not change the other answers, according to the documentation. The context window is 262,144 tokens across the shared input and all questions.
Developers can combine different media in that input, such as a call recording with account information or a scanned claim with damage photographs. Image and video resolution can be adjusted to control visual detail and token cost.
Pricing and performance claims
Google charges $0.10 per 1 million input tokens and $0 for output tokens, with a no-cost API tier in Google AI Studio for prototyping. The shared input is billed once per request, while each question adds its own instruction and criteria tokens.
The company says Gemini Decision 1 scores candidate options in a single forward pass rather than decoding answer text token by token. It reports median model-compute latency of 50–100 milliseconds and end-to-end latency of 150–250 milliseconds across multimodal inputs and contexts up to 16K tokens, with latency remaining nearly flat from 1 to 100 questions per request.
Across a shared benchmark suite, the Decisions API with Gemini Decision 1 was nearly five times cheaper than Gemini 3.5 Flash-Lite with the Interactions API, Google says. The company also claims its new model led the suite and had the lowest calibration error on tests covering agent routing, reasoning, ranking and multimodal understanding.
Use inside Google
Gemini Notebook uses the endpoint to route incoming queries to a model tier and thinking level, according to the company. Google says the routing serves everyday queries with up to 40% lower latency while reserving deep-reasoning models for complex tasks.
The company also says its Stitch rapid canvas experiment uses the API for product-flow wireframes and edits, while CC from Google Labs uses it to select emoji reactions from more than 200 options. Google Search uses the API to detect adversarial behavior and content against adversarial evaluations, the company says.
Integrations are rolling out across services including OpenRouter, LangChain and Vercel, alongside the google-genai SDKs; support for Gemini in Google Cloud is listed as coming soon. If an input safety filter blocks a request, the API returns an HTTP 400 error instead of answers, the documentation states.