Gemini Decisions API review: fast, cheap answers to yes/no and pick-one questions. Send the unsure ones to a person
A lot of what AI does inside business software isn't writing at all. It's deciding: is this ticket urgent, which team should get it, does this photo show damage. Google's new Decisions API is built only for that kind of question, and it changes the cost of asking it a thousand times a day.

Google's Gemini Decisions API, launched November 11, 2026, is a good fit for teams that make many bounded calls, such as routing tickets, checking a photo against a rule or grading content on a fixed scale. It answers yes/no, multiple-choice and ordered-score questions with a confidence number, at $0.10 per million input tokens and nothing for output. It doesn't write text, its benchmarks are Google's own, and it isn't in Google Cloud yet, so test it on your own cases and route low-confidence answers to a person.
What Google released
On November 11, Google introduced the Gemini Decisions API and a model built for it, Gemini Decision 1. It's available through the Gemini API in Google AI Studio.
It does one thing. You give it an input, which can be text, images, audio, video or PDFs, and a list of closed questions about that input. It answers each question by scoring the options you provided, without writing any text. There are three question types:
- Yes/no: the probability that a condition holds.
- Choice: pick one of 2 to 255 options, with a probability for each.
- Score: grade the input on an ordered scale you define, from lowest to highest.
Every answer comes with a confidence score between 0 and 1, and you can send up to 1,000 questions against one input in a single request. Google says the input can be up to 256,000 tokens.
Who it is for
This is a developer tool, so the buyer is a product or engineering team. But the use cases are very much business ones.
Think of the small decisions software makes all day: is this support ticket urgent, which team should get it, is this a refund request, does this return photo show damage, is this listing in the right category, does this message break our posting rules. Many teams answer those today by asking a general-purpose model and parsing the text it writes back. That works, but it's slower than it needs to be, costs output tokens and occasionally comes back in a format your code didn't expect.
A concrete case: an online store's returns desk receives a photo and a short note with each return. Three questions about that one input (does the photo show damage, which of five return reasons fits best, should this go to a person) come back in one call, each with a confidence. The confident ones move on; the rest land in front of someone who can look properly.
Google's own examples are of the same kind. It says Gemini Notebook uses the API to pick which model should handle each query, and Google Search uses it as a safety check on MCP tools.
Decision guide
A decision call or a generative call
| Consideration | Gemini Decisions API | A generative model asked to pick |
|---|---|---|
| Output | A choice or score from options you defined, plus confidence | Text you then have to parse |
| Many questions on one input | Up to 1,000 in one request, input billed once | Usually one prompt per question, or a long prompt to parse |
| Cost driver | Input tokens only ($0.10 per 1M) | Input plus output, sometimes thinking tokens |
| Explaining why | No written reasoning | Can explain, if you ask and pay for it |
| Open questions | Not possible; options must be fixed | Possible |
The model gives you a confidence number. You still have to decide what number is low enough to ask a person.
Pricing, as published
Gemini Decision 1 costs $0.10 per million input tokens, and output tokens are free because there aren't any. The shared input is billed once per request, however many questions you ask about it; each question adds only its own instructions and options. There's a no-cost tier in Google AI Studio for prototyping.
To put that in proportion with simple arithmetic: a support team that sends 10,000 tickets a day of about 500 tokens each, and asks four questions about every one, sends roughly 5 million input tokens a day, plus a little for each question's instructions. At the published price that's about fifty cents a day. Your real numbers will differ, but the order of magnitude explains why Google is pitching it for decisions you'd previously not have bothered to automate.
Google also claims speed: 150 to 250 milliseconds end to end, nearly flat from one to a hundred questions, and says the API is nearly five times cheaper than Gemini 3.5 Flash-Lite across its benchmark suite. Those are Google's numbers on Google's tests. They're plausible for this kind of model, but measure on your own traffic before you plan around them.
Where it falls short
It can't explain itself. You get a choice and a confidence, not a reason. For decisions you have to justify to a customer, a regulator or a colleague, you'll need another step.
Your options have to be right. It can only choose from what you give it. If your ticket categories are muddled, it will pick confidently among muddled categories. Writing clear options is now the main work, and it's the same discipline as setting an acceptance gate for a model's job.
The benchmarks are Google's. Google says it leads its benchmark suite with the lowest calibration error. Calibration, meaning whether "0.9 confident" is right nine times in ten, is exactly what you need to check on your own data, because your thresholds depend on it.
Not in Google Cloud yet. Google says support is "coming soon" to Gemini in Google Cloud. If your company buys AI through a Google Cloud contract with its own data terms, that matters. Integrations with OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI and others are rolling out.
Early. Google's sample code calls a beta API version. Plan for changes.
How the work moves
Route by confidence
- Ask
Urgent? Which team? Refund request? One call on the ticket text and any photo.
- Act when sure
Above your chosen confidence, route the ticket automatically.
- Ask a person when not
Below it, send to a triage queue with the model's top two guesses shown.
- Check weekly
Sample automatic routes and compare with what people would have done.
How it compares
Against a small generative model asked to classify, the trade is clear: the Decisions API should be faster and cheaper, and it can't drift into free text. You lose explanations and open-ended answers. If you're already sending easy work to a cheaper model, as in sending the cheap work to Haiku, this is the next step down for questions that have a fixed set of answers.
Against training your own classifier, it skips the training and the labelled data you'd need to start, and it handles images and audio in the same call. A custom model may still win when you have years of labelled examples and a stable set of categories.
The model gives you a confidence number. You still have to decide what number is low enough to ask a person. Start strict, send everything below it to a human queue with the model's top guesses shown, and loosen the threshold only after you've compared a few weeks of automatic decisions with what your team would have done.
Plain verdict
If your product makes lots of bounded decisions and you're paying a generative model to make them, the Gemini Decisions API is worth a trial this month. It's priced for volume, built for exactly this job and easy to put behind a confidence threshold. If your decisions need explanations, or you can only buy AI through Google Cloud, wait.
If you want help deciding which steps in your product should be decision calls, which need a generative model and where a person stays in the loop, that's our AI applications and agents work. Talk to us.
Related work: Spin:Market
We build alongside the people doing the work.
Bring us the problem you’re thinking about.