Quick Answer
Gemini Decision 1 (gemini-decision-1) is Google’s model for fast, bounded decisions, launched November 11, 2026 with the Gemini Decisions API. Instead of generating text token by token, it scores your candidate options in one forward pass and returns probabilities plus a confidence score (0 to 1) for each answer. You send one shared multimodal input of up to 256K tokens and up to 1,000 independent questions — yes/no checks, multiple choice, or rubric scores. It costs $0.10 per million input tokens with zero output tokens billed, and there is a free tier in Google AI Studio (200 requests per minute, 50,000 tokens per minute, 10,000 requests per day) for prototyping.
What Is Gemini Decision 1?
Gemini Decision 1 is a purpose-built model for one narrow job: making closed-set decisions. A normal Gemini model takes a prompt and writes an answer word by word; your code then has to parse that text back into a decision — did the user mean yes or no, which option won, how severe is this? Decision 1 skips the writing entirely. It looks at the input, scores every option you gave it, and hands back numbers your code can act on directly.
Google calls the Decisions API a “control plane” for AI apps: the layer that decides which tool to call, which model tier should answer, whether content passes a policy check, or how a piece of work should be graded. It is built on Gemini’s native multimodal architecture, so one shared input can mix text with images, audio, video or PDFs — for example, a call recording paired with account metadata, or a scanned PDF claim paired with damage photos.
It is a model, not a chatbot. There is no conversation, no explanation, and no generated prose. If you need a reason for a decision, an extracted field, or multi-step reasoning, you still need a generative model.
Release Date and Availability
Google announced the Gemini Decisions API and Gemini Decision 1 on November 11, 2026, in a Google DeepMind blog post by Min Choi and Felix Yu — one day after Gemini 4 Carbon went generally available. It is available now through the Gemini Developer API (v1beta) in Google AI Studio, with Python and JavaScript SDKs (google-genai 3.2 or later) plus REST.
Support for Gemini in Google Cloud is “coming soon” — for now it is the Gemini Developer API only. Ecosystem integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode. Google has not announced regional limits beyond the countries where the Gemini API is already offered.
Google is not the first to this idea. TypeSafe launched Jev in September 2026, OpenAI unveiled its own Decisions API at DevDay on September 29, 2026, and Microsoft followed with Microsoft-Decision-1, built on Alibaba’s Qwen.
Key Features
1. Answers up to 1,000 questions in one pass
One request can carry a single shared input and up to 1,000 independent questions about it. The input is billed once no matter how many questions you ask — each question only adds its own instruction and criteria tokens. Latency stays nearly flat from 1 to 100 questions per request, according to Google.
2. Three question types, mixed in one request
You can combine all three types against the same input:
- noul (yes/no): estimates the probability that a condition holds, with optional true/false description criteria. A yes/no probability is clamped to 0.01–0.99.
- choice (multiple choice): picks from 2 to 255 options, returning the winner and normalized probabilities for every candidate (they sum to 1).
- score (ordered rubric): grades against an ordered rubric starting at 0, returning a continuous weighted score and per-level probabilities.
3. Calibrated confidence on every answer
Every answer includes a confidence score from 0 to 1 — the probability that the top answer is right. Google’s guide tells you to set thresholds from labeled examples of your own traffic, weighing false positives against false negatives: act automatically above a threshold, and route ambiguous cases to a fallback or a human reviewer. Google reports the lowest calibration error across its benchmark suite, but that is a vendor-reported figure.
4. Native multimodal input, up to 256K tokens
One shared input can be text, images, audio, video or PDFs, up to 256K tokens (the context window is 262,144 tokens across the input and all questions). You can tune the resolution on image and video inputs to control visual detail and token cost. There are no separate audio transcription or OCR costs.
5. 150–250 ms end-to-end latency
Google reports median model compute of 50–100 ms and 150–250 ms end-to-end latency across multimodal inputs and contexts up to 16K tokens. That speed is the point: a decision that takes a quarter of a second can sit in front of every request in a live app.
6. Zero output tokens — a fixed, predictable bill
Because the model never generates text, there are no output tokens, no thinking tokens, and no variable reasoning bill. The cost of a request is fixed by its input. Google says the Decisions API is nearly 5x cheaper across its shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API — a vendor-reported comparison.
7. Already running inside Google products
Google says teams across the company are already using it: Gemini Notebook uses it as a real-time router, directing queries to the best model tier and thinking level, with up to 40% lower latency on everyday queries; Stitch’s rapid canvas experiment uses it to generate playable wireframes and apply voice or tap edits almost instantly; CC from Google Labs uses it to pick an emoji reaction from a palette of more than 200 options in a fraction of a second; and Google Search uses it as a safety layer for MCP tools to detect adversarial behavior.
Pricing Explained
Google lists one paid rate for the Decisions API: $0.10 per million input tokens, $0 for output tokens. The same rate applies across text, image, video, audio and PDF input.
| Tier | Price |
|---|---|
| Paid (Gemini API) | $0.10 per 1M input tokens |
| Output tokens | $0 (always 0 tokens) |
| Free tier (Google AI Studio, prototyping) | $0 — 200 RPM, 50,000 TPM, 10,000 requests/day |
Two billing rules make that rate go further than it looks. First, the shared input is billed once per request, however many questions you evaluate. Second, there is no thinking-token or output-token bill at all, so a verbose answer or a long reasoning trace can never multiply the cost the way it can with a generative model.
There are catches. There is no context caching, so a long context repeated on every call is billed in full each time. There is no Batch API or Flex tier, so you cannot halve the price for offline jobs the way you can with most Gemini and Claude models. Paid-tier rate limits rise with usage tiers: Tier 1 allows 1,500 requests per minute and 1M tokens per minute, Tier 2 allows 5,000 RPM and 4M TPM, and Tier 3 allows 15,000 RPM and 10M TPM.
How does it compare? Against Gemini 3.5 Flash-Lite ($0.30 input / $2.50 output per 1M tokens), Google’s own claim of nearly 5x cheaper lines up with independent list-price arithmetic on typical router and policy-check workloads. Against Claude Haiku 5.5 ($0.10 input / $0.50 output per 1M tokens), the gap is much smaller — Haiku matches the $0.10 input price, so Decision 1 wins mainly by billing zero output tokens and by charging the shared input once. See our What Is Claude Haiku 5.5? guide for that model. If cost is the only reason to switch, run your own numbers first — and note that Haiku’s 50% Batch discount would flip the math for offline work that can wait.
Gemini Decision 1 vs a Normal Gemini Model
| Feature | Normal Gemini model (e.g. Gemini 4 Carbon) | Gemini Decision 1 |
|---|---|---|
| Output | Writes text, token by token | No text — returns probabilities and confidence scores |
| What you get back | An answer to parse | Numbers your code can act on directly |
| Output tokens billed | Yes | $0 — always 0 output tokens |
| Thinking tokens billed | Yes, on reasoning models | None |
| Questions per request | One prompt at a time | Up to 1,000 independent questions on one shared input |
| Typical latency | Seconds, depending on length | 150–250 ms end-to-end (vendor-reported) |
| Best for | Writing, explaining, multi-step reasoning | Routing, triage, grading, policy checks |
| Conversation / tools | Supported | Not supported — no history or tool-call blocks |
The two are complements, not rivals. A common pattern is Decision 1 in front as the router — cheap, fast, with a confidence threshold — and a bigger generative model behind it for the cases that need real reasoning. If you want the wider Gemini family first, see our What Is Google Gemini and How Does It Work? explainer and our What Is Gemini 4 Carbon? overview.
How to Use Gemini Decision 1
1. Get an API key in Google AI Studio
Sign in at aistudio.google.com and create a Gemini API key. The free tier is available there for prototyping (200 requests per minute, 50,000 tokens per minute, 10,000 requests per day). Note that Google’s pricing page says free-tier data can be used to improve its products, while paid use is not.
2. Install the SDK (version 3.2 or later)
Use the google-genai SDK for Python or JavaScript, or call the REST endpoint POST /v1beta/decisions directly. The API is v1beta only for now.
3. Send one shared input and your questions
Build a request with client.decisions.create, model gemini-decision-1, one input (text, image, audio, video or PDF, up to 256K tokens) and a questions object. Google’s own example checks a photo of a breakroom counter plus a text note against all three question types in one request:
from google import genai
client = genai.Client(http_options={"api_version": "v1beta"})
result = client.decisions.create(
model="gemini-decision-1",
input=[{
"type": "image",
"uri": "https://storage.googleapis.com/generativeai-downloads/images/spill.jpg",
"mime_type": "image/jpeg",
},
{
"type": "text",
"text": "Breakroom counter check before the 9 AM client meeting.",
},
],
questions={
"has_spill": {
"type": "noul",
"instructions": "Is there an uncleaned liquid spill on the counter?",
},
"next_action": {
"type": "choice",
"instructions": "Which cleanup action fits the scene?",
"criteria": {
"wipe_with_cloth": "Small counter spill with a cloth or towel already nearby.",
"dispatch_janitorial": "Large floor hazard, broken glass, or active leak.",
"ready_for_meeting": "Counter is clean and ready.",
},
},
"severity": {
"type": "score",
"instructions": "Rate the severity of the counter mess.",
"criteria": [
"Clean and ready.",
"Minor localized spill.",
"Major spill or safety hazard.",
],
},
},
)
4. Read the answers and set thresholds
Each answer comes back with probabilities and a confidence score. Google’s guide tells you to set thresholds from labeled examples of your own traffic: act automatically when confidence is high, and fall back to a human reviewer or a bigger model when it is low. In the example above, the same photo used 1,326 input tokens (262 text + 1,064 image).
5. Try the demo apps first
Google points beginners to three demo apps in AI Studio — Doorstep Parcels, Speed Sketch, and Find the Zoo Keeper — and to the Python, JavaScript and REST guides in the Decisions API documentation.
Best Uses for Gemini Decision 1
- Model routing: classify each incoming query in milliseconds and send easy ones to a cheap model, hard ones to a deep reasoner. Gemini Notebook already does this, with up to 40% lower latency on everyday queries (vendor-reported).
- Support ticket and lead triage: sort messages by urgency, topic or department in one call instead of one generative call per ticket.
- Policy and content checks: run dozens or hundreds of rules over the same input — the input is billed once, so 20 rules cost little more than one.
- Grading and evaluation: score answers, essays or agent outputs against a rubric with the
scoretype, with per-level probabilities instead of a single unexplained mark. - Tool selection in agents: let the model pick which tool to call next from a fixed list, with a confidence threshold that escalates to a human when unsure.
- Image, audio and video triage: flag damage, defects or safety issues in media without paying for separate OCR or transcription.
- Live UI decisions: pick an emoji, a layout or a next step inside a product flow without adding noticeable latency — CC from Google Labs picks from 200+ emoji options this way.
Limitations and Things to Know
- It never explains itself. You get probabilities, not reasons. If a decision needs an explanation, an extracted field, or multi-step reasoning, use a generative model instead.
- Choice always picks something. Probabilities across options always add up to 1, even when nothing fits. Pair choice questions with a yes/no scope check, or threshold on confidence. Avoid keys like
none_of_the_above— Google advises using “other” with a clear description instead, because choice keys are sorted alphabetically and ties go to the first key. - Yes/no tops out at 0.99. A threshold above 0.99 never fires.
- Questions cannot see each other. Each question is independent. For a dependent decision, send a second request, or ask both and ignore one answer.
- Scores can hide splits. A score averages level indices, so a score of 1.12 can hide a split between levels. Read the per-level probabilities and keep scales linear.
- No conversation history or tool calls. Multi-turn context has to be serialized as plain text, and tool-call and thought blocks are rejected. Question instructions are text-only.
- No context caching, Batch API, or Flex/Priority inference. Repeated long contexts are billed in full every time, and there is no half-price batch option for offline work.
- Input is untrusted. Google warns that user text can argue for its own classification. Keep a human or a second check on payouts and other high-stakes actions.
- Google Cloud support is “coming soon”. For now it is the Gemini Developer API only, so Google Cloud-only shops should wait.
- Benchmarks are vendor-reported. Google says Decision 1 leads its benchmark suite (agent routing, reasoning, ranking, multimodal understanding) with the lowest calibration error, and that it is nearly 5x cheaper than Gemini 3.5 Flash-Lite — but Google did not publish per-benchmark scores or name the comparison models in the launch post, and no independent evaluation exists yet. Test calibration on your own labeled data before trusting it.
Frequently Asked Questions
What is Gemini Decision 1?
Google’s model for fast, bounded decisions, launched November 11, 2026. It scores up to 1,000 yes/no, choice or rubric questions about one shared input in a single pass, returning probabilities and confidence scores instead of text.
How is Gemini Decision 1 different from a normal Gemini model?
A normal Gemini model writes text token by token, and you pay for output tokens. Decision 1 never writes — it scores your options in one forward pass, bills zero output tokens, and answers in about 150–250 ms.
How much does Gemini Decision 1 cost?
$0.10 per million input tokens, $0 for output. There is a free tier in Google AI Studio: 200 requests per minute, 50,000 tokens per minute, 10,000 requests per day.
What is the model ID for Gemini Decision 1?
gemini-decision-1, used with the Decisions API (POST /v1beta/decisions or client.decisions.create in the google-genai SDK 3.2+).
How many questions can it answer at once?
Up to 1,000 independent questions per request, all against one shared input of up to 256K tokens. The input is billed once, no matter how many questions you ask.
What are the three question types?
noul (yes/no), choice (pick from 2–255 options), and score (grade against an ordered rubric). You can mix all three in one request.
Is Gemini Decision 1 free?
There is a no-cost tier in Google AI Studio for prototyping, capped at 200 requests per minute and 10,000 requests per day. Paid use starts at $0.10 per million input tokens.
Can Gemini Decision 1 explain its answers?
No. It returns probabilities and confidence scores, not reasons. Use a generative Gemini model when you need an explanation.
Final Verdict
Gemini Decision 1 is the most useful kind of launch: a narrow tool for a job most AI apps already do badly. Today, teams route, triage and grade with general-purpose models and then parse text back into decisions, paying for output tokens they never needed. Decision 1 returns probabilities instead, in a fraction of a second, with the input billed once and nothing for output. It is not a replacement for a generative model — it cannot explain, extract or reason step by step — but as the fast control plane in front of one, with a confidence threshold and a human fallback, it is hard to beat at $0.10 per million tokens. Start on the free tier, test its calibration on your own labeled data, and only then put it in front of real traffic.
Editorial update: November 12, 2026. This article is based on Google’s Decisions API launch announcement and the Gemini API documentation, checked November 12, 2026.