Gemini Decision 1 answers 1,000 questions in one pass and bills zero output tokens

A thousand questions, one forward pass, and not a single output token on the bill. On November 11, 2026, Google opened the Gemini Decisions API, backed by a new model, Gemini Decision 1, that doesn't generate text at all: it scores closed-set options and returns probabilities. It follows OpenAI and Jev with the same question format but, unlike them, it takes audio, video and PDFs.

01What it is and what it does

The point is to pull out of your generative LLM all the bounded decisions we currently handle with a prompt and a brittle parser: which model should get this query, is this ticket an outage, which tool should the agent call, does this text break a policy. Instead of writing an answer token by token, gemini-decision-1 scores the options you define in a single pass and returns a probability for each one.

There are three question types, and you can mix them in one request:

• noul (yes/no): the probability that a condition holds.
• choice: picks from 2 to 255 options and returns the winner plus the normalized distribution.
• score: grades against an ordered rubric (starting at 0) and returns a continuous weighted score.

Every answer also carries a confidence between 0 and 1, so your code can send ambiguous cases to a bigger model or to a human.

Google's numbers:

The detail that really moves the cost: the shared input is billed once per request, no matter how many questions you ask. Each question only adds the tokens of its own instructions and criteria. If today you run ten LLM calls over the same document, one per check, here you pay for the document once.

Google says it already runs this as a control plane: Gemini Notebook uses it to route each query to the right model and thinking level (up to 40% lower latency on everyday queries), and Google Search uses it as a safety layer for its MCP tools.

02Trying it out

You need an AI Studio key. Over REST it's a single POST to the v1beta/decisions endpoint:

curl -s https://generativelanguage.googleapis.com/v1beta/decisions \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-decision-1",
    "input": "Since 10:40 checkout returns 502 for every customer in the EU.",
    "questions": {
      "is_outage": {
        "type": "noul",
        "instructions": "Does the ticket describe a service outage?"
      },
      "team": {
        "type": "choice",
        "instructions": "Which team should handle it?",
        "criteria": {
          "billing": "Invoices, charges and refunds.",
          "platform": "Errors, latency or downtime.",
          "account": "Login, passwords and permissions."
        }
      }
    }
  }'

The response has an answers object with one entry per key: is_outage with its probability and confidence, and team with the winning option, the probabilities map and its confidence. In usage, total_output_tokens is always 0. From Python you need google-genai 3.2 or later (pip install -U google-genai), and the call is client.decisions.create(...) on a client set to api_version="v1beta".

03Gemini Decision 1 vs OpenAI Decisions and Jev

It isn't alone. OpenAI has its own Decisions API on gpt-6-luna, already generally available, and TypeSafe, the startup behind Jev, sells it through OpenRouter and other providers. Google even borrowed the vocabulary: noul is Jev's name for a yes/no question.

04Where it wins and where it loses

It wins on modalities and volume. It's the only one of the three that takes audio, video and PDFs directly. OpenAI rejects files and audio, and Jev asks you to turn audio and video into text first. If you're classifying call recordings, scanned invoices or clips, Gemini skips the transcription or OCR step, which isn't billed separately either. And 1,000 questions per request versus 400 matters when you apply a long policy, rule by rule, to the same document.

It wins on calibration, by its own numbers. Google claims the lowest calibration error on its benchmark suite. Simon Willison tried it on a video and saw the difference in style: Jev and OpenAI return 1.0 or 0.0 on yes/no questions, while Gemini answered, for instance, 0.92 to "is there a robot?" and 0.05 to "is there a dragon?". If you plan to escalate uncertain cases with thresholds, a probability that doesn't snap to 0 or 1 is exactly what you want.

It loses on price against Jev. At $0.042 per million, Jev costs less than half. For high-volume, text-only classification, that gap matters more than audio support you won't use.

It loses on maturity. OpenAI is already GA. Gemini Decision 1 lives in the Gemini Developer API (v1beta), and support in Gemini on Google Cloud is only announced as "coming soon". That's the catch that matters: if your company can only consume models through Vertex AI, for data residency or contract reasons, you can't put it in production yet.

One note on Google's own comparison: "nearly 5x cheaper" is against Gemini 3.5 Flash-Lite with the Interactions API on its benchmark suite, not against the current Gemini 4 models or the rivals in the table.

05Verdict

Who it's for: teams that already route requests between models, moderate content or check policies with a generative LLM and a parser, and especially anyone working with audio, video or PDFs. Moving those decisions to a model that only scores cuts latency and cost, and gives you a real probability to set thresholds on.

Who should wait: anyone who needs Vertex AI or a stable API, who today has OpenAI's GA option at the same price, and anyone classifying only text at very high volume, where Jev is cheaper. Either way, the question format is nearly identical across providers, so testing Gemini against your own cases is cheap.

06Sources

• Gemini Decisions API: Fast, calibrated multimodal decisions — Google, November 11, 2026.
• Decisions API — Google AI Studio docs.
• Decisions — OpenAI docs.
• What is Jev? — OpenRouter.
• Gemini Decisions API — Simon Willison, November 11, 2026.