Google has launched the Gemini Decisions API and a new model behind it, Gemini Decision 1 (gemini-decision-1). The service is built for bounded questions such as "is this a policy violation?", "which tool should the agent call?" or "rate this on a 0–4 rubric." It doesn't generate an answer token by token. Instead, Google says, the model scores the candidate options in one forward pass and returns probabilities with a confidence score from 0 to 1. Product manager Min Choi and principal scientist Felix Yu of Google DeepMind announced it on November 11. It's available now in the Gemini API and Google AI Studio.
This matters because so much of an agent's work is small decisions: routing, gating, grading and filtering. Running a full chat model for each one costs time and money, and parsing free text into a yes/no adds failure points. Decision models return a structured answer plus a number you can threshold. Low-confidence cases can then go to a bigger model or a human, the cascade pattern Hightouch tested with TypeSafe's Jev last week.
What it does
Google's post lists three question types:
| Type | What you get back |
|---|---|
| Yes/No | Probability that a condition holds, with optional criteria for "true" and "false" |
| Multiple choice | The winning option out of 2 to 255, plus normalized probabilities for each |
| Ordered rubric score | A weighted score on a rubric starting at 0, plus the probability of each level |
One request can carry up to 1,000 independent questions against a single shared input of up to 256K tokens. That input can be text, images, audio, video, PDFs or a mix. Image and video resolution can be turned down to save tokens.
Price and speed (Google's claims)
| Gemini Decision 1 | |
|---|---|
| Input | $0.10 per 1M tokens (shared input billed once, plus each question's text) |
| Output | $0, with no thinking-token, OCR or audio-transcription charges |
| End-to-end latency | 150–250 ms, for inputs up to 16K tokens |
| Median model compute | 50–100 ms |
| Scaling | Latency "nearly flat" from 1 to 100 questions per request |
| Free tier | Limited no-cost tier in AI Studio |
Google says the service is nearly 5x cheaper across its benchmark suite than running Gemini 3.5 Flash-Lite through its Interactions API. It also claims the lowest calibration error on benchmarks covering agent routing, reasoning, ranking and multimodal understanding. The benchmark figures in Google's post appear only in images. SiliconANGLE reports that Google says Decision 1 beat Jev and OpenAI's Decisions API on seven of nine benchmarks, while Jev came out slightly more cost-efficient. These are all Google's own tests, with no independent results yet.
Where Google already uses it
Google's examples come from its own products. Gemini Notebook uses the model as a real-time router to pick the model tier and thinking level, and Google says everyday queries got up to 40% lower latency. Google Search uses it as a safety layer for MCP tools to spot adversarial behaviour. A Google Labs experiment uses it to pick an emoji reaction from more than 200 options. No outside customers are named.
The competition
According to SiliconANGLE, TypeSafe's Jev launched in mid-September, and TypeSafe then raised $870 million at a $7.5 billion valuation. OpenAI followed with its own Decisions API in October. Google is the third entrant, and its pitch is price (output is free) plus native multimodal input in one call.
Availability and limits
- Now: Gemini API and Google AI Studio, with Python, JavaScript and REST guides. Google's sample code calls the
v1betaAPI version. - Rolling out: integrations with OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode.
- Coming soon: Gemini in Google Cloud, via a sign-up form. The post doesn't mention Vertex AI regions or a GA date.
- Scope: only the three question types. It won't write, summarize or reason out loud, so it complements a general model rather than replacing one.
Who should care
Teams running agents or moderation pipelines should look at it first, especially those currently prompting a small LLM to answer "yes" or "no" and parsing the reply. One practical approach is to replay a labelled sample of your own routing or policy decisions through Decision 1, Jev and OpenAI's endpoint. Then compare accuracy at the confidence threshold you'd actually use, not just headline benchmark wins. Calibration claims only matter if the scores hold up on your data.