Google launches Gemini Decisions API: up to 1000 checks in a single request

An agent receives a customer inquiry: should it respond directly, look up information, or hand the task to a human operator? To make this choice, developers often call a standard language model, wait for a response, and parse the resulting JSON. Google has introduced Gemini Decisions API, a dedicated service that evaluates predefined options without generating text sequentially. According to the company, latency is 150–250 ms, and pricing is $0.10 per million input tokens. For teams, this is a reason to check how much time and money their products spend on intermediate decisions.

One document, many checks

The API runs on the new Gemini Decision 1 model. Developers provide shared inputs and a set of independent questions. The service supports text, images, audio, video, and PDFs: for example, a call recording can be combined with account details, or a scanned claim with photos of a damaged product.

Consider an online store's returns service. A customer sends a photo of a broken mug and writes that the packaging was intact. A single request can check for visible damage, select the next step from the available options, and rate the severity of the problem on a predefined scale. The shared materials are submitted once, and answers come back with probabilities.

Three formats are available: yes-or-no checks, selection from 2–255 options, and ratings on an ordered scale. In the last case, the service returns a weighted score and probabilities for individual levels. A request can contain up to 1000 questions, with an input context of up to 256 thousand tokens.

The practical benefit is less repeated work. If a document needs to be checked against several rules, it does not have to be submitted in a separate generative call for each rule. The questions must be independent, however: a chain in which the next step depends on the previous answer still needs to be orchestrated in code.

What the speed and pricing figures mean

Google reports 50–100 ms for model computation itself and 150–250 ms of end-to-end latency. But the measurement conditions matter: these figures apply to contexts of up to 16 thousand tokens. The company reports almost unchanged processing time as the number of questions increases from one to a hundred. There is currently no basis for extrapolating those figures to the maximum 1000 questions and the full available context.

The shared input is billed once per request. Each question adds tokens for its own instructions and criteria. Output tokens are not billed, and there are no separately billed reasoning tokens either. At the stated rate, processing 10 thousand input tokens would cost $0.001, provided that is the request's entire billable volume.

The company also says that, on its benchmark suite, the service is almost five times cheaper than Gemini 3.5 Flash-Lite through the Interactions API. That is the result of a specific comparison. In your own application, you will need to account for instruction length, the mix of files, and the number of repeat calls. For images and video, the level of detail can be adjusted, affecting token volume and cost.

Probabilities help select a fallback

The key practical feature is the probabilities attached to answers. For example, a team can process straightforward returns automatically and send ambiguous ones to a human operator. Confidence thresholds need to be tuned using your own examples: Google's claim of good calibration does not yet establish how reliably the service will assess photos of your particular products.

There is also a limitation in how the task is framed. If the correct action is missing from the options provided, picking a winner will not solve the problem. Production workflows should include an option for human handoff or additional review. Safety checks also need separate testing: Google mentions using the API to protect MCP tools in Search, but the supplied announcement contains no detailed metrics with which to assess that use case.

You can get started through Google AI Studio; a free tier for prototypes and guides for Python, JavaScript, and REST have been announced. A sensible approach is to take an existing routing or classification task and compare quality, latency, and cost on real data. If intermediate calls account for a significant share of an agent's runtime, that experiment will provide a clear answer about the new API's value.