Google launches Decisions API: routing and classification for $0.10 per million tokens

Almost every LLM product has a boring layer that nobody shows in demos: a router, a classifier, a filter. It decides which model to call, whether to pass a request through a policy check, and whether the retrieved context is relevant. Usually teams take a small generative model, ask it to answer with a single word, and then parse the result. Google proposes getting rid of this hack: the Gemini Decisions API was built specifically for selection tasks.

What it is

It is a dedicated endpoint built on the Gemini Decision 1 model. It does not write text. In a single pass it evaluates all the options and returns probabilities. There are three kinds of questions:

• yes or no, with the probability that the condition is met;
• a choice among 2–255 options, with a probability for each;
• a score on a scale starting from zero, with a continuous value.

The input can be up to 256 thousand tokens: text, images, audio, video, or PDF. For a single piece of content, up to a thousand independent questions can be asked in one request.

What it looks like in practice

Imagine a support desk that receives tickets with attached screenshots. Previously you made a separate call for each attribute: is it a refund or not, is the customer irritated or not, which of twelve teams to route it to, how urgent it is. Now the ticket is sent once, along with a list of questions. In return you get a set of probabilities and scores that you can compare directly against thresholds in your code.

This approach also suits agents. Instead of asking a large model to "pick a tool and explain," you get a distribution over tools. If confidence falls below a threshold, the request goes to a more expensive model or to a human. This is ordinary engineering logic, except now it relies on numbers rather than on parsing a string.

Cost and calibration

Under the pricing, a million input tokens cost 10 cents, and output is not charged. The shared content of a request is paid for once, no matter how many questions are asked about it. So a long document with a hundred questions costs the same as the same document with one.

Google claims that on its own test suite covering agent routing, reasoning, ranking, and multimodal tasks, the model shows the lowest calibration error. Calibration here means the honesty of probabilities: if the model says "80%," it should be right in roughly four cases out of five. For thresholds and escalation this matters more than raw accuracy. The same materials claim nearly fivefold savings compared with Gemini 3.5 Flash-Lite via the Interactions API.

What to keep in mind

First, all the figures so far come from Google's own benchmarks. Before replacing a working router, run the model on your own labeled data and check calibration yourself: it determines whether you can trust the thresholds.

Second, access is limited for now. The API is available through the Gemini API and Google AI Studio, with Google Cloud support promised later. If you work within an enterprise Vertex environment, you will have to wait. AI Studio offers a free tier for prototypes.

Third, the ecosystem is catching up gradually. Integrations with OpenRouter, LangChain, Vercel, LiteLLM, and other tools do not appear all at once, so rough edges are possible at launch.

What to do

Find the places in your stack where an LLM is used as a switch: routing between models, moderation, assessing context relevance in RAG, tool selection by an agent. These are the first candidates to test with the Decisions API. If the probabilities turn out to be honest, you will get a cheaper and more predictable control layer, while expensive generative models are reserved for cases where you actually need text.