Google on Nov. 11 introduced the Gemini Decisions API, a new service that answers up to 1,000 closed-ended questions — yes or no, multiple choice or scoring per rubric — in a single pass over text, images, audio, video or PDF, and returns probabilities and confidence scores in 150 to 250 milliseconds. It costs $0.10 per million input tokens and doesn't charge output tokens.

Unlike generative models, which produce the response on a token-by-token basis, the new Gemini Decision 1 model scores candidate options in a single calculation step. The developer sends a multimodal input of up to 256,000 tokens—for example, a call recording along with account data, or a scanned PDF with photos of damage—and that input is charged only once, no matter how many questions are evaluated in the request.

The API offers three types of decision in the same request: yes/no, which estimates the probability that a condition will be met; multiple choice, which chooses between 2 and 255 alternatives and returns the normalized probabilities of each; and rubric score, which rates the input on an ordered scale. Each response includes a confidence score of 0 to 1, so that the code refers doubtful cases to a backup system or a human reviewer.

According to Google, Gemini Decision 1 leads in agent routing, reasoning, ranking, and multimodal understanding testing, with the lowest calibration error of the models compared, and is nearly five times cheaper than Gemini 3.5 Flash-Lite with the Interactions API in that same test suite. The model's median latency is 50 to 100 milliseconds and remains nearly flat at 1 to 100 questions per request.

Google already uses it indoors: Gemini Notebook uses it as a real-time router to route each query to the right model, with up to 40% less latency in everyday queries and reserving deep reasoning models for complex tasks; Stitch generates playable wireframes in seconds and applies edits by voice or touch; CC, from Google Labs, chooses the reaction emoji from more than 200 options; and Google Search uses it as a security layer to detect adversarial behavior in MCP tools.

For teams building agents in Latin America, arithmetic is what changes the game: A routing or validation decision no longer requires paying for an entire generation at a time. At $0.10 per million tokens, a typical check of a few thousand tokens costs a fraction of a cent, allowing guardrails, policy checks, and tool selection to be put into each agent step without the cost skyrocketing. It's the kind of primitive that makes orchestrators viable with many small agents, such as those running on their own servers.

The API is available today via the Gemini API in Google AI Studio, with a free tier for prototypes, and with the google-genai SDKs. There are already integrations with OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode; support for Gemini within Google Cloud is coming soon.

The release comes in the same week that Google expanded access to Gemini 4 Carbon, a model designed for long and complex tasks: code, cybersecurity, financial research, and coordination between various agents. Decisions is the opposite extreme: quick and cheap individual evaluations. A word of warning: the test results are from Google and have not been independently verified, and the model has just been announced, so its actual performance outside the lab remains to be seen.

The next step will be to see how it is adopted by agent frameworks and if Google brings it to Google Cloud, where enterprise deployments live. If price and latency are sustained in production, the cheap and calibrated decision can become the standard piece between the model that reasons and the one that acts.

Sources and credits: Google Blog — Gemini Decisions API: Fast multimodal decisions, by Min Choi and Felix Yu · ETCIO / ANI — Google Launches Gemini Decisions API for Rapid, Cost-Effective AI Decision-Making

Cover: Google.

Originally published in Digitalia.