Google Launches Gemini Decision 1, a Model That Answers 1,000 Questions in One Pass
Google DeepMind launched the Gemini Decisions API on November 11, 2026, powered by a new model called Gemini Decision 1. Unlike a chatbot, the model writes no text at all. It scores up to 1,000 closed questions about a single piece of input in one pass and returns a probability and a confidence score for each answer.
A model built for judgments, not conversation
Most large language models answer by generating text one token at a time. Gemini Decision 1 works differently. A developer sends one multimodal input, text, images, audio, video, or PDFs, up to 256,000 tokens, along with as many as 1,000 independent questions. Instead of decoding an answer, the model scores the candidate options in a single forward pass and returns structured numbers that software can act on directly.
The API supports three question types. A yes-or-no check returns the probability that a condition holds. A multiple-choice question picks from 2 to 255 named options and reports the probability of each. A rubric question returns a weighted score across a scale the developer defines. Every answer carries a confidence score, and output token usage is always zero because no text is generated.
Google positions the model for the small, repetitive judgments inside agent workflows: choosing which tool to call, checking a policy, routing a request to the right model, or grading a piece of context. The company says Gemini Notebook already uses the model to route queries to different models and levels of reasoning effort, with up to 40 percent lower latency for routine queries in that workflow.
Speed and price
Google reports end-to-end latency of 150 to 250 milliseconds, with median model compute of 50 to 100 milliseconds. That figure was measured on inputs of up to 16,000 tokens and with up to 100 questions per request, so teams using the full 256,000-token window or the full 1,000-question limit should measure their own workloads.
Pricing is 0.10 dollars per 1 million input tokens, with no charge for output tokens. A no-cost tier in Google AI Studio lets developers prototype, with free-tier limits of 200 requests and 50,000 input tokens per minute. Google says the model paired with the Decisions API is nearly five times cheaper than Gemini 3.5 Flash-Lite through the Interactions API on the shared benchmark suite the company used for comparison.
The model is available now through the Gemini Developer API at the v1beta endpoint and in Google AI Studio. Support for Gemini in Google Cloud is listed as coming soon.
How it stacks up
Decision models are a new category, and Google is joining a field that formed quickly. TypeSafe launched the first such model, Jev, in mid-September. OpenAI made its own Decisions API generally available on October 22, running on GPT-6 Luna, with a limit of 400 questions per request and the same base price of 0.10 dollars per million input tokens.
In Google’s own launch benchmarks, Gemini Decision 1 led five of the seven tests that all four compared models ran. It scored 92.2 percent on JevBench, ahead of GPT-6 Luna at 90.9 percent and Jev 1.14 at 90.0 percent. Jev led on two tests: the browser-task benchmark Mind2Web Multimodal, 86.8 percent to 86.2 percent, and When2Call, which tests tool-use decisions, 82.0 percent to 79.9 percent.
Google also reports the lowest calibration error in its comparison, at 0.0250, against 0.0494 for Jev 1.14, 0.0721 for GPT-6 Luna, and 0.2104 for Flash-Lite. Calibration measures how closely a model’s confidence matches how often it is actually right, which matters when a confidence threshold decides how many cases reach a human reviewer.
The independent Artificial Analysis Decision Index v2.1 offers a more modest picture. It scores Gemini Decision 1 at 65.98 and ranks it ninth of 18 decision models, behind Jev 1.14 at 66.23. TypeSafe also undercuts Google on price for text-only work, charging 0.042 dollars per million input tokens for Jev 1.14.
Why it matters
Agents spend much of their time and budget on small decisions: which tool fits this request, whether a document meets a policy, how urgent a case is. Running a full generative model for each of those checks is slow and expensive, and its free-text answers then have to be parsed back into structured data. A purpose-built decision model removes both steps, returning a number the application can branch on immediately.
The trade-off is scope. Gemini Decision 1 cannot explain itself, extract a value from a file, or draft a reply. Those jobs still need a generative model, and questions that depend on each other have to be handled in application code or with follow-up calls. The model is a specialist for closed questions, not a replacement for general models.
For teams building agents, the practical move is to pilot it on a checklist workflow where one item arrives in several formats, such as a support ticket with a recording and an attachment, or a claim with a form and photos. Run it in parallel with the existing process first, set a confidence threshold for escalation to a human, and measure latency and cost on real traffic before letting its output change customer records.
