Gemini Decisions API: closed questions, numeric answers, and how it compares with Jev

When your software needs a small decision, such as “Is this message a refund request?”, you would usually ask a chat model. Then you read its response and turn it into a decision. Google launched the Gemini Decisions API on 11 November 2026 to shorten that process. Send a question and get a number your software can act on directly. Not tried yet.

What it is

Send one input: text, an image, audio, video or a PDF. Include up to 1,000 questions. Each question is one of three types:

• Yes or no (Google calls it noul): returns the probability that the answer is yes, between 0.01 and 0.99.
• Choice: picks from a list of 2 to 255 options, with a probability for each.
• Score: places the input on a scale whose levels you define, from lowest to highest.

Each answer includes a confidence score between 0 and 1. In your code, you decide when the software acts on its own and when it sends a case to a person. Google says requests usually take 150 to 250 milliseconds with inputs of up to 16,000 tokens and 1 to 100 questions. The price is $0.10 per million input tokens. Output is free. The input is counted once, regardless of the number of questions.

How it compares with Jev

TypeSafe AI released its Jev model on 15 September 2026, about two months before Gemini. Simon Willison wrote that Gemini is almost identical. It has the same three question types, including the name noul. Here are the differences:

• What it reads: Jev reads text and JSON only. Gemini also reads images, audio, video and PDFs.
• Price: Jev costs $0.042 per million tokens, less than half Gemini’s price. Output is free on both.
• Access: Jev is in early access with a waitlist. Gemini is open to everyone, with a free trial in AI Studio.
• Probability of yes: In Simon Willison’s test Jev gave yes/no questions a probability of 1.0 or 0.0. Gemini stays between 0.01 and 0.99, and gave 0.92, for example, to “Is there a robot in the video?”.

In October 2026, OpenAI launched a similar tool with the same three question types. It reads text and images and costs the same as Gemini.

If all you have is text, Jev is cheaper. For audio recordings, video and PDFs, Gemini is the only one of the three that reads them.

How to start

You need a Google account and Python 3 on your computer. Allow about 15 minutes by my estimate (not tried yet). The trial is free in AI Studio.

1. Open aistudio.google.com/api-keys and create an API key. Don’t put it in any file you share.

2. Install the library (version 3.2 or later). The second line waits for you to paste your key and press Enter; the key will stay hidden on screen:

   pip install -U google-genai
   read -rs GEMINI_API_KEY && export GEMINI_API_KEY

   TerminalCopy

3. Create the example file from Google’s docs and run it:

   cat > decide.py <<'EOF'
   from google import genai
   
   client = genai.Client(http_options={"api_version": "v1beta"})
   
   result = client.decisions.create(
       model="gemini-decision-1",
       input="The blue mug I ordered on Monday arrived cracked on Wednesday.",
       questions={
           "is_damaged": {
               "type": "noul",
               "instructions": "Did the item arrive damaged?",
           },
       },
   )
   
   answer = result.answers["is_damaged"]
   print(f"P(yes)={answer.noul:.2f} confidence={answer.confidence:.2f}")
   EOF
   python3 decide.py

   TerminalCopy

   Check: you should see a line like P(yes)=0.99 confidence=0.99. If you get an error about the key, run the second line of step 2 again in the same Terminal where you run the file.

4. Ask the coding agent you use to adapt the file to your data:

   Read Google's Decisions API docs at https://aistudio.google.com/docs/decisions-api. Adapt decide.py to my case: <what the input is, for example customer emails>. Ask these questions: <your yes/no, choice or score questions, with their options or levels>. Send every case with confidence below <threshold, for example 0.7> to a list for a person to review. Compare numbers and dates in code, not inside the questions. Do not put my API key in the file.
   

   PromptCopy

   A good result is a file that prints an answer and confidence score for each input, with unclear cases set aside.

If it doesn’t work as expected:

• Error 400: The error message names the rule you broke. Question names can contain only English letters, digits, _ and -.
• An option that doesn’t fit wins: A choice question always distributes probability across your options, even when none fits. Add an “other” option with a clear description, and check the confidence score.
• Mistakes with numbers and dates: Google recommends comparing them in code, then including the result in the input.

Limits

• It returns a number without an explanation. Writing about Jev, Simon Willison warned that a number alone could hide bias. He said he hoped nobody would use it to rank job applicants. Gemini also returns a number.
• Accuracy and speed figures are Google’s own; Jev’s figures are TypeSafe’s own.
• Google has not mentioned Arabic. Arabic is not tested.
• It’s available through the Gemini API only for now; Google Cloud support is “coming soon”.
• User-written text may try to influence the answer. Google recommends clear criteria and human review for important decisions such as payments.