Gemini Decision 1: 10-Cent AI That Answers Yes or No in 200ms

Google launched Gemini Decision 1 on 11 November 2026: it answers yes or no instead of writing text. The new Gemini Decisions API takes one shared input — text, images, audio, video or a PDF, up to 256,000 tokens — plus up to 1,000 typed questions, and returns a calibrated probability and confidence score for each one. Google says it takes 150 to 250 milliseconds and costs $0.10 per 1M input tokens, with no output charge.

That matters because most of what an AI agent does all day is not writing. It is choosing: which tool to call, whether a message is safe, whether an answer is good enough. Until now every choice cost a full model call, a generated reply, and a parser. Decision 1 replaces that with a scored answer in about a fifth of a second.

What Gemini Decision 1 actually is

Gemini Decision 1 is a new Google model, model ID gemini-decision-1, built for one job: closed-set decisions. You give it an input and a list of candidate options — yes or no, a choice from a short list, or a score against a rubric — and it returns how likely each option is. Google's launch post, written by Min Choi and published on 11 November 2026, calls it a model for fast, calibrated multimodal decisions. Calibrated means a 0.9 is more trustworthy than a 0.6, so you can set a threshold and act on it. Read the full announcement on Google's blog. Our chatbot-versus-agent guide explains why deciding is the expensive part of an agent.

The model runs behind a new endpoint, the Gemini Decisions API, at POST https://generativelanguage.googleapis.com/v1beta/decisions on the Gemini Developer API. It is available through the Gemini API in Google AI Studio, with Vertex AI support coming soon. You need google-genai 3.2 or later, and there is a no-cost tier in AI Studio.

How the Decisions API works

The request shape is simple. You send one shared multimodal input and up to 1,000 typed questions about it: yes/no, a choice between options, or rubric scoring, such as "is this document a valid invoice?". The model returns a calibrated probability and a confidence score for each question. You get numbers back, not sentences.

Because the input is shared, asking 100 questions about the same document costs little more time than asking one. Google says latency stays nearly flat from 1 to 100 questions per request, at 150 to 250 milliseconds end to end, of which 50 to 100 milliseconds is model compute. That flat curve makes one call enough for grading and triage.

The numbers: 200 milliseconds and 10 cents

Both headline figures are Google's claims. The latency figure, 150 to 250 milliseconds end to end, is fast enough to sit inside a live conversation without the user noticing. Our how to build an AI voice agent guide and our AI calling agent cost guide both stress that latency is a cost in voice work.

The price is $0.10 per 1M input tokens, with $0 output tokens because the output is always zero tokens. Compare that with the flagship end of the Gemini family: our Gemini 4 Argon pricing guide puts that model at $2 per 1M input tokens until 30 November 2026, then $4. Decision 1 costs a twentieth of the introductory Argon input price. At that price you can check everything, not just the important cases.

What a business can use it to decide

Google names four uses in the launch post: tool selection, policy checks, model routing and context grading. Tool selection is choosing which tool to call. Policy checks are the safety gate: is this request allowed? Model routing sends easy questions to a cheap model and hard ones to a flagship. Context grading judges whether retrieved documents are relevant before the answer is written.

The business uses beyond Google's list are easy to see. Content moderation triage: score every incoming comment for abuse and hold only the borderline ones for a human. RAG answer grading: score every generated answer for grounding before it reaches the customer, the quality gate our RAG chatbot guide recommends. Ticket routing, which our AI agents for small business automation guide lists first, plus lead scoring and document checks, all become one-call jobs. Our WhatsApp chatbot cost guide shows how small per-message costs add up at volume.

What Decision 1 cannot do

The limits are as important as the capabilities. Decision 1 does not generate text. If you need a summary, an explanation, a draft reply or any writing at all, this model cannot do it — writing summaries or explanations is unsupported. It only scores the options you give it, so it cannot replace your main model.

There are platform limits too. Batch, caching, Flex and Priority are unavailable on this endpoint, so no discounted overnight batch scoring. Vertex AI support is listed as coming soon, so use the Gemini Developer API in AI Studio for now. One honest caveat: Google's launch post shows benchmark results only as images, with no published scores to quote, so do not claim Decision 1 leads any benchmark. Our GDPx agent benchmark post showed even the best models score only around 12% on long-horizon agent tasks. Judge Decision 1 on your own tests, as our Meituan LongCat-2.5 coverage did with unverified figures.

The judge-and-worker pattern

The useful way to think about Decision 1 is as a judge in front of a worker. The worker is a full model — for example Gemini 4 Hydrogen, the Gemini 4 flagship our Gemini 4 Carbon coverage describes — doing the expensive work. The judge is Decision 1, deciding which tool to use, whether the answer is good enough, and whether the request is safe.

This pairing is the practical way to make agents fast and affordable.

What the judge checks

The judge checks tool choice, safety, and answer quality: the three decisions an agent makes most often.

Today, many agents burn a full language-model call plus parsing every time they must choose anything, and those calls add up. A cheap, fast judge that answers in 200 milliseconds replaces most of those calls. Our Qwen4-Max Preview coverage made the same point about the system around the model. Our n8n vs Make vs Zapier comparison shows the same routing step inside a workflow, and our AI agents and automation development service builds systems like this.

How to try it this week

Start in Google AI Studio, where the no-cost tier lets you prototype without a bill. You will need google-genai version 3.2 or later. Pick one decision your system already makes badly or expensively and write the questions as a typed list. Use a real input and record the probability and confidence for each question.

Then set a threshold and test it. Send a hundred real examples through and check where the model is confident, wrong, or uncertain. Send the uncertain band to a person. A decision model earns its place when you can trust its confident answers and catch its uncertain ones, so measure both before you wire it into anything live. If you want help designing that split for an AI calling agent or a voice receptionist (see our AI receptionist guide), talk to Coheser. Our AI automation for logistics guide shows how the same pattern works in a different industry, and our custom software cost guide covers build costs. More from Coheser on AI agents and automation.

Frequently asked questions

What is Gemini Decision 1?

Gemini Decision 1 is a Google model, model ID gemini-decision-1, launched on 11 November 2026. It answers closed-set questions — yes/no, choices and rubric scores — returning a calibrated probability and confidence score for each, instead of generating text.

How fast is the Gemini Decisions API, and what does it cost?

Google says 150 to 250 milliseconds end to end, nearly flat from 1 to 100 questions per request. The price is $0.10 per 1M input tokens with $0 output tokens, and there is a no-cost tier in Google AI Studio for prototyping. Both figures are Google's claims.

Can Gemini Decision 1 write summaries or explanations?

No. It does not generate text at all. Writing summaries or explanations is unsupported. Use it to decide — which tool, which route, is this safe, is this answer good. Use a full model for the writing.