AI and security briefsModels

What is the Gemini Decisions API, and should it replace my routing and policy prompts?

By Brandon Da Costa, founder of Digiton Dynamics

Published 12 November 2026Updated 12 November 2026Editorial policy

Google released the Gemini Decisions API on 11 November 2026. It scores up to 1,000 yes or no, choice and rubric questions in one call and bills no output tokens. Test it on the steps where you pay a generative model to produce a label, and keep a person on anything high stakes.

Who should act: Teams that call a generative model only to pick a route, a tool, a queue or a yes or no, and care about the latency and cost of that step.

What changed

Google introduced the Gemini Decisions API on 11 November 2026, powered by a new model, Gemini Decision 1 (gemini-decision-1), through the Gemini API in Google AI Studio (Google blog, 11 November). It writes no text. It scores your candidate answers in one pass and returns probabilities plus a confidence value from 0 to 1. One request takes a shared input of up to 262,144 tokens across text, images, audio, video and PDFs, and up to 1,000 independent questions (Google AI Studio docs, 12 November).

There are three question types. Yes or no returns a probability between 0.01 and 0.99. Choice picks from 2 to 255 options and returns a probability for each. Score grades against 2 to 255 ordered rubric levels.

Google lists $0.10 per million input tokens and no output charge, with a no-cost tier in AI Studio for prototyping. The shared input is billed once per request and each question adds only its own instruction and criteria. Google claims 150 to 250 ms end to end, and says the setup is nearly 5x cheaper than Gemini 3.5 Flash-Lite on the Interactions API across its own benchmarks (Google blog, 11 November). Those are Google's figures.

The REST path is on v1beta and needs SDK 3.2 or later. Google says integrations are rolling out in OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, and that Gemini in Google Cloud support is coming soon.

Who it affects

• Teams that route requests between model tiers with an LLM call. Google says Gemini Notebook uses the API as its router, with up to 40% lower latency. That is Google's own number.
• Teams that check inputs or agent output against many policy rules. One call can hold many rules, and Google says Search uses it as a safety layer for MCP tools.
• Anyone who needs generated text, summaries or JSON in their own schema. Google's docs point those cases to the Interactions API and structured output.
• Teams that can only buy through Google Cloud. Support there is listed as coming soon.

What we checked

Run on 12 November 2026 on the Mac that builds this site. Each command and its output are copied as they ran.

$ curl -s https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/ | grep -o -E '\$0\.10 per 1M input tokens|150-250 ms end-to-end' | sort -u
$0.10 per 1M input tokens
150-250 ms end-to-end

$ curl -s https://aistudio.google.com/docs/models/gemini-decision-1 | grep -o -E 'Input token limit\*\* [0-9,]+'
Input token limit** 262,144

The price, the latency claim and the input limit are on Google's pages as of today. Google's benchmark claims were not rerun.

What to do now

1. List every call whose output is only a label: a route, a queue, a tool name, a yes or no flag. Leave generation alone.
2. Pull a few hundred real inputs from your logs and write the correct label for each. Google says to set thresholds from labeled examples of your own traffic.
3. Run your current method and the Decisions API over the same set. Compare accuracy at several confidence cutoffs, latency and input tokens. The call shape from Google's quickstart:

   curl -X POST "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":"MESSAGE","questions":{"queue":{"type":"choice","instructions":"Which queue handles this?","criteria":{"billing":"Payment questions.","other":"Anything else."}}}}'

4. Add a catch-all key such as other, because a choice question spreads probability across your options even when none fits. Docs advise a separate yes or no scope check too.
5. Send low-confidence answers to a fallback or a person. Treat input as untrusted, since user text can argue for its own label, and keep a second check on payouts and other high-stakes actions.

What to watch

• Independent calibration results. Every benchmark so far is Google's.
• Changes to the v1beta path. Pin the SDK version.
• The pricing and rate limit pages for the tier you would use, which Google keeps separate.
• Availability in Gemini in Google Cloud.

Questions people ask

Digiton's production review checks which model each agent step calls, what it costs per accepted result and who reviews low-confidence cases, and the checks are on the vibe coded app audit page.

How this brief was made: researched from the sources below, drafted with AI assistance and checked against each source on the dates shown. See the editorial policy.

Sources

1. Google blog, Gemini Decisions API: Fast, calibrated multimodal decisions, 11 November 2026, read 12 November 2026. https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/
2. Google AI Studio documentation, Decisions API, read 12 November 2026. https://aistudio.google.com/docs/decisions-api
3. Google AI Studio documentation, Gemini Decision model page, read 12 November 2026. https://aistudio.google.com/docs/models/gemini-decision-1

More briefs

How these briefs are sourced and corrected. All AI and security briefs