Gemini Decisions API: what it is and what it costs

The Gemini Decisions API is a new Google endpoint that answers fixed questions with probabilities instead of writing text: yes or no, pick one of these options, or grade this against a rubric. It costs $0.10 per million input tokens, output tokens are free, and it runs on a new model called Gemini Decision 1.

Google launched it on Wednesday. @sundarpichai described it as "built for closed-set choices," scoring up to 1,000 questions in a single pass. @OfficialLoganK, who leads the Gemini API, called it "our new frontier decision model in the Gemini API" and said you can try it for free. The @Google account added the speed claim: answers with confidence scores in 150 to 250 milliseconds end to end.

Perplexity shipped the same idea in October, and we covered it in Perplexity Decisions API: what it is and what it costs. Google's version handles more questions per request and more kinds of input, and it costs more.

How the Gemini Decisions API works

You send one shared input and a set of named questions. Per Google's docs, each question has one of three types:

• noul is yes/no. You get back the probability that the answer is yes.
• choice picks from 2 to 255 options and returns the winner plus a probability for every option.
• score grades against an ordered rubric and returns a weighted score. Google allows up to 255 levels and recommends 2 to 10.

Every answer also carries a confidence number from 0 to 1, which Google defines as the probability that the top answer is right. I think that number is the most useful part, because you can let the model handle everything above, say, 0.9 and send the rest to a person.

The input can be text, images, audio, video or PDFs, and everything has to fit in a 262,144-token window. Google's launch post says its own teams already use it: to route everyday queries in Gemini Notebook, to pick emoji in CC from Google Labs, and as a safety check on MCP tools in Search.

Gemini Decisions API pricing vs Perplexity

The Perplexity column is the launch pricing from our October post. Its CEO said then that the price would drop further, so check the current number before you compare.

Google bills the shared input once per request, however many questions you attach. Each question adds only its own instructions. So 10,000 support emails at about 500 tokens each comes to roughly 50 cents, plus a little for the questions. There are no thinking-token charges and no separate fee for transcribing audio or reading a PDF.

The docs also say this endpoint has no context caching and no Batch API, so if you send the same input twice you pay for it twice.

How to try it

Google says there is a no-cost tier in AI Studio for prototyping. The quickstart is short:

1. Get an API key in Google AI Studio.
2. Install the SDK: pip install -U google-genai (the docs call for version 3.2 or later).
3. Create a client with api_version set to v1beta.
4. Call client.decisions.create with the model gemini-decision-1, your input, and your questions.

The docs example passes a one-line customer complaint about a mug that arrived cracked and asks a single noul question, "Did the item arrive damaged?" The response comes back as 0.99 for yes with 0.99 confidence. You could add "which department handles this" as a choice question and "how upset is the customer" as a score question to the same request without paying for the complaint text again.

If you would rather click than code, AI Studio has a Decisions Playground and a few demo apps. The model is also listed on OpenRouter, and Google names LangChain, the Vercel AI SDK and LiteLLM among the launch integrations. Google Cloud support is "coming soon." If you are already on Google's newest chat model, our guide on how to use Gemini 4 Carbon covers that side.

My read

Google did not publish benchmark numbers in the text of its post, only charts and a claim that the model "leads across the suite." I would hold that loosely until someone outside Google runs it against Perplexity's model and TypeSafe's Jev, the decision model that started this category in September.

Google's clear advantage is input. Perplexity reads text and images, while Gemini Decision 1 can take a voicemail recording or a scanned PDF invoice and answer questions about it directly. For a business where the messy input is a call or a document, that could matter more than the price gap.

What you give up is the exit. Perplexity's weights are open, so you can run them yourself if the terms change, and Google's are closed.

What this means for a small business

Most office automation is a chain of small calls, like whether a voicemail is an emergency or whether an invoice is a duplicate. A decision model makes each of those cheap and predictable enough to run on every message, and the confidence score tells you which ones still need a human.

Before trusting it, run a few hundred of your own past decisions through it and see where it disagrees with what your team did. If you are not sure which decisions in your week are worth handing off, New Face Design's free process audit maps where the hours go.