Google launched the Gemini Decisions API on November 11. It runs on a new model, Gemini Decision 1 (gemini-decision-1), and it does one job well: it answers closed questions like "which one?", "is this true?", or "how strong is this on a scale?" and tells you how confident it is. A single request can carry up to 1,000 questions. It comes back in about 0.2 seconds, and the output is free.
How it differs from a regular LLM
When you ask an AI agent to judge something, like "which tool should handle this message" or "does this image break the rules," the usual approach is to have a large language model write out a full answer and then dig the verdict out of the text. That's slow, costs more than it should, and sometimes answers a different question than the one you asked.
The Decisions API works differently. You list all the options up front, and the model works out a probability for each option in one pass instead of generating words one at a time. That's why it's fast. Every answer also comes with a confidence score between 0 and 1, so your code can say "send this to a human if confidence drops below a threshold I pick."
Three kinds of questions
| Type | What to ask | What you get back |
|---|---|---|
| Yes or no | Does this image break the rules? Is this email a payment reminder? | The probability that the statement is true |
| Pick one | Which model should handle this request? Which tool should run? | The winning option and the probability for every option, up to 255 options |
| Scale rating | Is this exam answer poor, fair, or good? | A continuous score and the probability for each level |
You can mix all three types in one request, and they share a single input. The input can mix text with pictures, sound, video clips or PDFs, as long as it stays under 256K tokens. For example, you could send a customer service recording along with the customer's file, or a scanned insurance claim form plus a few photos from the scene.
Speed and price
- Fast: 150 to 250 milliseconds end to end. Asking one question or a hundred takes about the same time.
- Cheap: $0.10 per million input tokens, and output is free. You pay for that shared input a single time in each call, so each extra question only adds the tokens for its own short instructions.
- Free tier: Grab an API key from Google AI Studio and you can try it for nothing.
The chart comes from Google's official blog. On the shared test set, Google says it costs nearly 5 times less than calling Gemini 3.5 Flash-Lite through the normal API, and its calibration error (how far the confidence scores drift from how often the answers are actually right) is the lowest of the group.
What Google uses it for
- Routing questions: Gemini Notebook uses it to decide which model tier each question should go to and whether it needs extended thinking. Wait times for everyday questions drop by up to 40%.
- Picking emoji: Google Labs' CC uses it to instantly pick the right emoji out of more than 200 options.
- Safety checks: Google Search uses it to spot malicious behavior aimed at MCP tools.
What it means if you build AI agents
If you're building an AI agent, this fits best at the steps where the agent has to make a small call over and over: which tool to run here, whether this result needs a second check, whether this request should go to the cheap model or the expensive one. Those calls used to be hardcoded rules, or a full LLM call every time. Now they can go to something that's fast and nearly free.
Support is landing in tools such as OpenRouter, LangChain, LiteLLM, Vercel, OpenCode and Pydantic AI, and Gemini on Google Cloud is coming soon.
Want to try it first? Send this to your AI agent:
Help me try Google's new Gemini Decisions API (model gemini-decision-1).
1. First check whether my computer has a Gemini API key in the environment variable GEMINI_API_KEY. If it doesn't, stop and tell me where to get a free key in Google AI Studio. Don't sign up for me.
2. Using the official Decisions API docs, write a minimal Python example. The input is [a piece of text, for example three customer emails]. Ask three questions:
- Yes or no: is this email a complaint?
- Pick one: which team should handle it, [sales / after-sales / finance]?
- Scale: rate the urgency as [not urgent / normal / very urgent].
3. Show me each answer and its confidence score in a table, and flag any answer with confidence below 0.7.
4. Tell me how many input tokens this request used and roughly what it cost.
Watch out
- It doesn't write text. Summaries, rewrites, and open-ended answers still need a regular model. This one only picks from the options you give it.
- You have to define the options clearly. Overlapping or vague options spread the probability thin and push confidence scores down.
- The free tier has limits. Check the quotas in AI Studio before you launch anything real.
- The speed and "5x cheaper" numbers come from Google's own tests.
Comments (0)
No comments yet. Start the conversation.