Not every model call needs a sentence back. A router deciding which agent should handle a query, a guardrail checking a policy, a grader scoring retrieved context: each wants one answer and a confidence number. Generating a paragraph and parsing it is slow, expensive and fragile. Gemini Decision 1, released on 11 November 2026 as the model behind Google's new Decisions API, skips the paragraph entirely.
Answers without text
The model scores options in a single forward pass rather than generating answer text. That is why it produces zero output tokens. Every answer comes with a confidence score between 0 and 1, so your code can send ambiguous cases to a fallback model or a human.
Three question types are supported:
- Yes/No: the probability that a condition holds, with optional criteria describing what counts as true or false.
- Multiple choice: a winner from 2 to 255 options, plus normalised probabilities for all of them.
- Ordered rubric: a grade against a rubric starting at zero, returned as a continuous weighted score with per-level probabilities.
One call can carry up to 1,000 questions against a single shared input. Ask whether a support ticket is urgent, which team owns it and how angry the customer sounds, all at once.
Limits and latency
| Field |
Value |
| API ID |
gemini-decision-1 |
| Released |
11 November 2026 |
| Shared input |
Text, image, audio, video, PDF, up to 256K tokens (262,144) |
| Output |
Decisions and confidence scores; no text |
| Questions per call |
Up to 1,000 |
| Latency |
150 to 250 ms end to end, for contexts up to 16K tokens |
| Price |
$0.10 per 1M input tokens; output not billed |
Median model compute is 50 to 100 ms, and Google says latency stays nearly flat from one question to a hundred. Image and video resolution can be turned down to trade detail for token cost.
Google describes it as launched without attaching a GA or preview label. Assume the interface may still evolve and pin your SDK version.
Billing that ignores output
Input costs $0.10 per million tokens. The shared input is billed once per request, however many questions you attach; each question adds only its own instruction and criteria tokens. There are no thinking-token charges and no separate fees for audio transcription or OCR. Caching and Batch are not offered. Input on the free tier is free. Google claims it comes out nearly five times cheaper than Gemini 3.5 Flash Lite across its shared benchmark suite. That is Google's claim, on Google's suite; treat all prices here as indicative.
Where it earns its place
Google names tool selection, policy checks, model routing and context grading as the targets. Inside Google, it routes queries in Gemini Notebook and sits in a safety layer that detects adversarial behaviour in MCP tools for Search. Integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, OpenCode and the google-genai SDKs. Support for Gemini in Google Cloud is listed as coming soon.
It is not a general model. It cannot explain itself, write a reply or call tools. Pair it with a generative model such as Gemini 3.8 Flash for anything that needs words, and let Decision 1 pick when to call it.
Calibrating your thresholds
Calibration is the claim to test. Label five hundred real decisions from your logs. Run them through, bucket the answers by confidence, and check that the 0.9 bucket really is right about nine times in ten. Then set your fallback threshold where the error rate becomes unacceptable for that decision. A routing call can tolerate more mistakes than a safety check.
Similar closed-set models are grouped under structured decisions, and the wider catalogue is on the Google provider page.