Google Says Its New Gemini Model Scores 1,000 Questions at Once Without Writing a Word

Google has built a Gemini model that refuses to chat. Gemini Decision 1, the engine behind the new Gemini Decisions API announced on November 11, does not write answers. You hand it one input and a stack of closed questions, and it hands back probabilities and confidence scores your software can act on. No prose, no output bill.

Google says it costs $0.10 per million input tokens, nothing for output, and scores up to 1,000 independent questions against one shared input in a single pass. It is also late. Google is not inventing this category. It is the latest big name to follow a startup into it.

Google DeepMind product manager Min Choi and principal scientist Felix Yu pitch the API for "bounded choices like tool selection, policy checks, model routing, and context grading." That is the dull plumbing inside an AI agent: which tool runs next, does this message break a rule, is this query easy enough for a cheap model.

"Rather than generating answer text token by token, Gemini Decision 1 scores your candidate options in a single forward pass, returning probabilities and confidence scores." Min Choi and Felix Yu, Google DeepMind

The model takes text, images, audio, video and PDFs, up to 256,000 tokens per input. Questions come in three types: yes/no (Google calls it noul), a choice between 2 and 255 options, and a score against an ordered rubric. Each answer carries a confidence score from 0 to 1, which Google says lets code send unclear cases to a fallback or a human reviewer.

The numbers are all Google

Every performance figure in the launch is Google's own. The company says end-to-end latency runs 150 to 250 milliseconds and stays nearly flat from 1 to 100 questions per request, but it ties that to contexts up to 16,000 tokens, not the full window. It says the setup is "nearly 5x cheaper" than Gemini 3.5 Flash-Lite across a shared benchmark suite, and that Gemini Decision 1 leads that suite "with the lowest calibration error." No outside results back any of it yet.

Google says it already runs the API in-house. Gemini Notebook uses it to route queries to the right model tier, which Google claims serves everyday queries "with up to 40% lower latency," a Notebook figure rather than a promise to developers. Google says Search uses it as a safety layer for MCP tools and Stitch uses it to build playable wireframes.

Access is narrower than the Gemini brand implies. The API is live in Google AI Studio with a no-cost prototyping tier. Support for Gemini in Google Cloud is only "coming soon," with no date, so enterprises buying through Google Cloud wait.

The Decoder called this Google joining a "decision-model wave," noting that after startup Jev introduced its version, OpenAI, Cloudflare and Microsoft followed. Developer Simon Willison was blunter.

"This new model API is almost identical to Jev: it supports the same three types of questions (yes/no, choice, and rubric-scoring) and even reuses Jev's term 'Noul' for a yes/no question." Simon Willison, developer and blogger

In his same-day note, Willison says Gemini Decision 1 costs the same as OpenAI's GPT-6 Luna decision model, and that Jev "remains the cheapest of the three, at 4.2 cents per million." So the 5x boast is measured against Google's own chat models, not the rivals fighting for this job.

Google's real edge, by Willison's account, is range. He says Jev handles text and JSON only and OpenAI handles text and images, while Gemini covers audio, video and PDFs too. In his quick video test, Gemini returned softer probabilities, such as 0.92 for "robot," where he says Jev and OpenAI give flat 1.0 or 0.0 answers. For anyone routing real traffic, a model that admits doubt may beat one that sounds sure.

After big generative launches like the paid rollout of Gemini 4 Carbon, this is the opposite: a small model whose whole job is yes, no or which one. The open question is whether Google's multimodal reach beats a startup that got there first and still charges less.