Google’s launch artwork for the Gemini Decisions API, November 11, 2026. Image: Google
Google has released a Gemini model that never writes a word. Gemini Decision 1, the engine behind the new Gemini Decisions API launched on Wednesday, reads text, images, audio, video or PDFs and answers up to 1,000 yes/no, multiple-choice or rubric questions about them in one go, handing back probabilities in about a quarter of a second.
How a model answers without writing
A chatbot builds its answer one token at a time. Gemini Decision 1 instead scores a fixed set of options in a single pass over the input, which is why Google pitches it for the quick, repetitive judgments inside AI agents: picking which tool to use, checking a request against policy rules, sending a query to the right model, or grading an answer.
Developers send one shared input of up to 262,144 tokens, alongside a list of questions. Google’s documentation describes three types:
- Yes/no: the probability that a statement holds, between 0.01 and 0.99.
- Choice: the winning option from 2 to 255 candidates, with a probability for each.
- Score: a weighted grade against an ordered rubric, with the probability of each level.
Every answer comes with a confidence score from 0 to 1, so an app can hand unclear cases to a person or a bigger model. Google says the model needs 50 to 100 milliseconds of compute, 150 to 250 milliseconds end to end, and stays nearly as fast whether it is asked one question or 100. Inputs can be mixed, such as a call recording with account details, or a scanned insurance claim with photos of the damage.
$0.10 per million tokens, and nothing for the answers
Gemini Decision 1 costs $0.10 per million input tokens, with no output tokens to pay for. The shared input is billed once per request however many questions are asked; each question adds only its own instructions. Google says there are no separate charges for transcribing audio or reading text in images, and that it works out nearly five times cheaper across its test suite than running the same jobs on Gemini 3.5 Flash-Lite.
It is available now through the Gemini API and Google AI Studio, which has a free tier for prototyping. Google’s pricing page says free-tier data can be used to improve its products, while paid use isn’t; the free tier is capped at 200 requests a minute and 10,000 a day. Python, JavaScript and REST guides are out, the SDK calls go to Google’s beta endpoint, and Batch API and caching aren’t supported. Support in Gemini in Google Cloud is “coming soon”, with integrations rolling out to OpenRouter, LangChain, Vercel, LiteLLM and others. Google hasn’t announced regional limits beyond the countries where the Gemini API is already offered.
Google’s scorecard beats OpenAI, but not on everything
Google is late to this kind of model. TypeSafe launched Jev in September, OpenAI unveiled its own Decisions API at DevDay on September 29, and Microsoft followed with Microsoft-Decision-1, built on Alibaba’s Qwen. So Google’s chart compares Gemini Decision 1 directly with Jev 1.14 and with OpenAI’s GPT-6 Luna, the only model behind OpenAI’s Decisions API. These are Google’s own figures and haven’t been independently checked:
| Google’s test | Gemini Decision 1 | TypeSafe Jev 1.14 | OpenAI GPT-6 Luna |
|---|---|---|---|
| JevBench (TypeSafe’s own benchmark) | 92.2% | 90.0% | 90.9% |
| New Yorker caption contest | 79.7% | 72.0% | 68.4% |
| When2Call (when to use a tool) | 79.9% | 82.0% | 70.4% |
| Calibration error (lower is better) | 0.025 | 0.049 | 0.072 |
| Cost of seven shared tests | $0.88 | $0.68 | $0.93 |
Gemini Decision 1 tops seven of Google’s nine tests, but Jev wins two, including web-browsing decisions, and comes out cheapest. Calibration matters here: a well-calibrated model that says 90% is right about nine times in ten, so software can trust its confidence. The audio and video tests have no Jev or OpenAI score because neither accepts those inputs. OpenAI’s documentation now lists its Decisions API as generally available, at the same $0.10 per million input tokens, but it takes only text, images and JSON, and caps requests at 400 questions to Google’s 1,000.
Already running inside Google
Google says its own teams use the model as a fast control layer. Gemini Notebook uses it to send each query to the right model and thinking level, which Google says cuts waiting by up to 40% on everyday questions. The Stitch design tool uses it to build clickable wireframes in seconds, Google Labs’ CC uses it to pick emoji reactions from more than 200, and Google Search uses it as a safety check on MCP tools, the connectors that let AI agents reach other services. It arrives a day after Google opened Gemini 4 Carbon to paying users.
Why it matters
AI agents make thousands of small yes-or-no calls, and running each one through a full chatbot is slow and expensive. With Google joining OpenAI, Microsoft and TypeSafe, decision models are becoming a standard part of the toolkit, and Google’s audio, video and PDF support gives it the widest range of inputs for now. The benchmark lead is Google’s own claim until others test it.
Sources: Google, Gemini Decisions API: Fast, calibrated multimodal decisions; Google AI Studio docs: Decisions API, Gemini Decision, pricing, rate limits; OpenAI, Decisions API guide.
Latest Google news
