Not tested by us yet. The model is on Google's own Gemini API, which our test harness does not call yet, and it is not on OpenRouter at the time of writing. Everything below comes from Google's announcement of the Gemini Decisions API. We'll update this post when we have run it on our own classification set.
Gemini Decision 1 (gemini-decision-1) is a Google model that never writes text. You send one input and up to 1,000 typed questions, and it returns a probability and a confidence value for each answer, in a single forward pass. It costs $0.10 per million input tokens, and output tokens are free. It is built for the small judgment calls inside agents: picking a tool, checking a policy, routing a query, grading retrieved context.
| Gemini Decision 1 | |
|---|---|
| Maker | |
| Announced | 2026-11-11 |
| API model ID | gemini-decision-1 |
| Inputs | Text, images, audio, video, PDFs, mixed in one shared input, up to 256K tokens |
| Question types | noul (yes/no), choice (2 to 255 options), score (ordered rubric starting at 0) |
| Questions per request | Up to 1,000 against the same input |
| Output | Probabilities plus a confidence from 0 to 1 |
| Price | $0.10 per 1M input tokens, $0 output |
| Access | Gemini API and Google AI Studio (free tier for prototyping). Google Cloud "coming soon" |
What a decision model is
A chat model answers by generating tokens that you then parse. A decision model skips generation and scores the candidate answers you supply. Google describes Gemini Decision 1 as scoring your options "in a single forward pass", so there is no JSON that fails to parse and no thinking tokens to pay for.
The request is a list of content blocks, in the same format as Google's Interactions API, plus a dictionary of named questions. Each question has a type, instructions and, for choice and score, criteria. The shared input is billed once per request, whatever the number of questions. Each extra question adds only its own instruction and criteria tokens. Google also says audio transcription and OCR are not billed separately.
This is the same family of model as Kakao's open Kanana-2-v-decide and Salesforce's Koa Decisions, with the same noul, choice and score question types.
What is new
Three things stand out compared with the earlier decision models we covered.
- Many questions, one input. Scoring 1,000 questions against a single document or video is a different cost structure from 1,000 separate LLM calls. Google says latency is "nearly flat" from 1 to 100 questions per request, for contexts up to 16K tokens.
- Any modality in one request. Text, images, audio, video and PDFs share one input, with image and video resolution adjustable to trade detail for token cost.
- Confidence as a routing signal. Every answer carries a 0 to 1 confidence, so code can send uncertain cases to a fallback model or a human.
Integrations with OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode are described as rolling out, so a hosted route we can test may appear soon.
The claims, and who made them
Every number below is Google's, none is ours.
- Latency: median model compute of 50 to 100 ms, 150 to 250 ms end to end. Google states this for contexts up to 16K tokens, while inputs can reach 256K.
- Accuracy: it "leads across the suite" on Google's own benchmarks for agent routing, reasoning, ranking and multimodal understanding, with the lowest calibration error. The results are published as charts only, so there are no numbers to quote.
- Cost: nearly 5 times cheaper across the same suite than Gemini 3.5 Flash-Lite through the Interactions API.
- In production at Google: Gemini Notebook uses it to pick a model tier and thinking level, with up to 40% lower latency on everyday queries. Google Search uses it to detect adversarial behavior in MCP tools.
The benchmark suite is Google's, the comparison set is not listed in the text, and a calibration claim means little until someone checks it on labels they own. Calibration is the point of a decision model: a 0.9 confidence should be right about 90% of the time.
Who should care
- Agent builders paying for routing and guardrail calls. If a Flash-class call currently decides which tool to run, a $0.10 per million input price with free output changes the cost of that step, if the accuracy holds. Our guide to cheap-first routing with escalation shows the generative version of that job.
- Teams classifying one long document many ways. A single PDF or call recording scored against hundreds of questions is where the shared-input billing should pay off most.
- Anyone on the Gemini API already. There is no new vendor, key or data agreement to sign.
Skip it if you need an explanation for each decision, since it returns a distribution and not reasoning. Also skip it if your data has to stay on your own hardware: this one is hosted only, unlike the open decision models above.
What we will test when we run it
We plan to put it next to TypeSafe's Jev, an open decision model if one passes its health check, and a cheap and a mid-priced chat model acting as classifiers, on one set of 30 to 60 items. We will include control items with obvious answers and ask each yes/no question in both directions, because a model that flips its answer when the wording is mirrored cannot be trusted for routing. Cost per 1,000 decisions, median latency and calibration go in the same table. Until then, treat the 150 to 250 ms figure as a vendor claim. For where decision models fit among the other recent releases, see the Koa Decisions write-up.