Gemini Decisions API: what it is, when to use it

On 11 November Google shipped the Gemini Decisions API, powered by a new model called Gemini Decision 1. It does not write a reply. It scores options.

That is the useful part. Most of the cost in an LLM product is not the one thoughtful answer. It is the dozens of small judgements around it: what kind of message this is, which model should handle it, whether a reply passed a check. A dedicated scorer is the right tool for those jobs, if the scores hold up.

Key takeaways

• The API answers yes/no, multiple-choice and rubric questions in one pass, with a confidence score on each.
• Google prices it at $0.10 per 1 million input tokens and bills no output tokens.
• Use it for routing, policy checks and grading. Do not use it to talk to a person.
• Low confidence should go to a fallback or a human, not to a silent default.
• Treat a swap onto it like any model change: evals first.

What the Gemini Decisions API actually does

You send one input, which can mix text, images, audio, video and PDFs, and up to 1,000 independent questions about that input. Google says the context can run to 256K tokens. Instead of decoding an answer token by token, Gemini Decision 1 scores your options in a single forward pass.

Three question types:

1. Yes/no (Google’s docs call this noul): the probability that a condition holds.
2. Choice: 2 to 255 options, with a winner and a probability for each.
3. Score: an ordered rubric, with a weighted score and per-level probabilities.

Every answer includes a confidence from 0 to 1. That is the knob that matters. A classifier that is quietly 0.51 sure is how you ship a wrong route.

Google quotes 150 to 250 ms end-to-end, and says latency stays nearly flat from 1 to 100 questions on contexts up to 16K tokens. Measure it on your own traffic before you trust those numbers.

What it costs, in their numbers

Google lists $0.10 per 1 million input tokens and $0 for output. The shared input is billed once even if you ask a hundred questions. Each question adds only its own instruction tokens.

They also say Decision 1 was nearly 5x cheaper than Gemini 3.5 Flash-Lite on their shared benchmark suite, because you are not paying for generated tokens. Take that as a vendor claim. The pricing page is the part you can check.

There is a no-cost tier in Google AI Studio. OpenRouter, LangChain, LiteLLM and a few others are listed as rolling out support.

Where it actually fits

Routing. “Is this a hard turn or a cheap labelling job?” is a closed choice. A lot of teams send that question to a small generative model and parse the prose. A scorer that returns probabilities in a couple of hundred milliseconds is a cleaner fit. Google says Gemini Notebook already uses this API to pick a model tier, and claims up to 40% lower latency on everyday queries. Put it behind the same config switch you use for providers, so you can turn it off.

Pass/fail grading. Offline evals are mostly binary checks on real conversations. Scoring those checks in one batch, with a confidence, is closer to an eval harness than a chat API. Ambiguous cases go back to a person or to the lead model.

Policy that is actually a list. “Does this message ask to store a new fact?” “Is this a title request?” Those are yes/no questions you can write down. If confidence is low, do not guess. Fall back.

This is the same split as routing between LLM providers: the user-facing reply and the hidden judgements are different jobs, and they should not share a bill.

Where it does not

The conversation itself. The Decisions API does not write. Putting a person’s question through a rubric and calling that “help” is the wrong product.

Anything you cannot undo. A crisis path, a payment, a delete, a send-to-production. Confidence scores are not a substitute for a person. If the score is used at all, it should only decide to stop and ask, never to proceed.

A replacement for evals. A fast classifier will be wrong in narrow ways. Do not move live traffic onto it because the demo looks cheap. Build a fixed set of cases, compare it with the current cheap model, and keep the old path until the new one wins on the checks that matter. Same rule as when a cheaper model shows up.

An API you cannot leave. Routing belongs in configuration. If Decision 1 sits behind the same interface as your other providers, you can try it. If the product starts to depend on Google-only question types with no fallback, that is a lock.

How to try it without fooling yourself

Write the questions as if a junior colleague had to apply them. If two people would disagree on the label, the model will too.

Log the confidence. Pick a threshold before you look at production traffic, then see how much would have been sent to a fallback. If most of the savings disappear once you honour low confidence, the API is not saving you money. It is hiding misses.

Keep the generative model on the path the user reads. Use the scorer on the path they never see.

We build this kind of routing and eval layer as LLM app development, and we review it as an LLM audit. If you want that put in your repo, write to us.

FAQ

What is the Gemini Decisions API?

A Google API, shipped 11 November 2026, that scores closed questions against one shared input. You send text, images, audio, video or a PDF plus up to 1,000 yes/no, multiple-choice or rubric questions. It returns probabilities and a confidence score, not generated prose.

How much does the Gemini Decisions API cost?

Google lists Gemini Decision 1 at $0.10 per 1 million input tokens and $0 for output tokens. The shared input is billed once per request. There is a no-cost tier in Google AI Studio for prototyping.

Is the Gemini Decisions API for helping a person decide?

No. It classifies and scores. A person still has to choose. Use it to route a request or grade a check, not to replace a conversation.

Should I replace my chat model with it?

Only for the bounded jobs: intent, policy, routing, pass/fail grading. Anything that needs a written reply still needs a generative model. Run your evals before you switch either path.