Route and check inputs with Gemini Decision 1, no LLM needed

Gemini Decision 1 answers yes/no, choice and score questions about text, images, audio, PDFs and video with a confidence score and zero output tokens, so your code can route, gate and audit automatically.

Prompt Engineering · YouTube · published Nov 13, 2026 · Automation & Agents

Route and check inputs with Gemini Decision 1, no LLM needed — video by Prompt Engineering▶ Watch on YouTube
Prompt
Using the Gemini Decisions API (model gemini-decision-1), check this customer message with three questions: did the item arrive damaged (yes/no), which support queue fits (choice, include an 'other' option), and how frustrated is the customer (score). Only act automatically if confidence is above my cutoff; otherwise escalate to a person.

Visual breakdown

Gemini Decision 1 is a decision model: it never writes text, it answers yes/no, choice or score questions about text, images, audio, PDFs and video, and returns a probability plus a confidence score. Your code compares that to a cutoff and acts, or escalates to a human or a bigger model.

The problem: Using a full LLM for small yes/no decisions means paying for thinking and output tokens, waiting for them, and then parsing the text that comes back.

The decision-model pattern: one input, a few questions, an if statement

  1. InputText, image, audio, PDF or video, even mixed in one call
  2. Prepare questionsSmall yes/no, choice or score questions in plain English, each option described
  3. Gemini Decision 1Scores every option in one pass, zero output tokens
  4. Confidence cutoffYour code checks probability and confidence against a threshold you set
    Dropped: Below cutoff goes to escalation
  5. EscalateA person or a bigger model handles the unsure cases
  6. Automatic actionCode runs the next step, for example approve a claim or click the next button
InputsYour codeAI modelResult

Three kinds of questions

  1. Yes/no: returns the probability that the answer is yes
  2. Choice: you give a list of options, get the winner plus a probability for every option
  3. Score: you give an ordered rubric, get a number on that scale

By the numbers

0output tokens per call
$0.10per million input tokens (paid tier)
~$10per million routing calls
~$170same job on a Flash model

Try it yourself

The free tier of Google AI Studio is enough for every example in the video, and the code is in a Colab linked in the description.

  1. Get an API key from Google AI Studio
  2. Create a client and call decisions create with the model gemini-decision-1, an input and your questions
  3. Give each question a type (yes/no, choice or score) and instructions in plain English
  4. Swap the input for an image, audio clip, PDF or video; the rest of the code stays the same
  5. Set a confidence cutoff in code; act above it, escalate below it

Examples built in the video: a browser agent that completes a checkout, an expense-claims auditor, and a voice agent.

Mistakes he made, and the rules that came out of them

  • Always give a choice question a way out such as 'other': without it, 'what's the weather in Paris' was routed to technical support with 92% confidence.
  • Send only what the question needs and compare numbers in code: pasting the whole expense policy made a date violation leak into the tip answer.
  • Ask small questions one at a time: one big 'what next' question split the confidence between valid options and the agent stopped.
  • Describe exactly what you mean, such as what 'empty' means for a card field showing grey example digits.
  • A decision model can't explain its answer; use a normal LLM when you need explanations or multi-step reasoning, and test on your own data, not only synthetic data.

Distilled from the video’s captions

How to try it

  1. Open Gemini.
  2. Copy the prompt and fill in the [placeholders].
  3. Watch Prompt Engineering's video for the full walkthrough and setup.

Related use cases

Share