Google Launches Gemini Decision 1: a $0.10-per-Million-Token Model That Scores 1,000 Questions Without Generating a Word

Google on Tuesday launched Gemini Decision 1, its first purpose-built decision model, along with a Gemini Decisions API now live in Google AI Studio. CEO Sundar Pichai announced the release on X, framing it as a model designed for "decision tasks over a closed set of options" rather than open-ended text generation. Instead of decoding an answer token by token, the model scores pre-defined candidates in a single forward pass and returns the choice, a probability, and a 0-to-1 confidence score.

The API supports three decision types that can be mixed within one request: binary yes/no judgments with the probability that a condition holds; multiple choice over 2 to 255 options, with normalized probabilities returned for every candidate; and ordered rubric scoring, where developers define grade levels and the model returns a continuous weighted score. Inputs can be text, images, audio, video or PDF, with a 256K-token multimodal context window, and a single request can evaluate up to 1,000 independent questions — one demo asks the model to judge a spill on an office counter, pick a cleanup procedure, and rate severity simultaneously.

Speed and price are the pitch. Google says median compute latency stays at 50-100 milliseconds on contexts up to 16K tokens, end-to-end latency runs roughly 150-250 milliseconds, and that latency stays nearly flat as the question count in a request scales from 1 to 100. Pricing is set at $0.10 per million input tokens with zero output-token billing; on Google's own benchmark set the company says calling the Decisions API costs nearly 5x less than invoking Gemini 3.5 Flash-Lite through the Interactions API.

Google published a scorecard against three competitors — OpenAI's GPT-6 Luna, the decision-specialist TypeSafe Jev 1.14, and its own Gemini 3.5 Flash-Lite — claiming first place on 7 of 9 benchmarks, all company-reported and not independently verified. On JevBench it reports 92.2% accuracy versus 90.9% for GPT-6 Luna, 90.0% for Jev 1.14 and 86.6% for Flash-Lite. It also leads Amazon ESCI product-search relevance at 58.7 Macro-F1, New Yorker Caption at 79.7 versus Luna's 68.4, PubMedQA at 79.8, and Global-MMLU-Lite at 86.0 versus 79.6. Calibration is the headline claim: a Top-1 Expected Calibration Error of 0.0250, roughly 65% lower than Luna's 0.0721 and 88% lower than Flash-Lite's 0.2104. Notably, Jev 1.14 still wins the two agent-flavored tests — Mind2Web Multimodal (86.8 vs 86.2) and When2Call tool-calling judgment (82.0 vs 79.9).

Google is already using the model internally. In Gemini Notebook, the Decisions API acts as a real-time router that picks model tiers and thinking depth per request, which Google says cuts daily-query latency by up to 40%. Google Stitch's new Rapid Canvas experiment uses it to produce interactive product-flow wireframes in seconds, Google Labs' CC app uses it to pick emoji reactions from more than 200 options, and Google Search uses it as a safety-check layer for MCP tool calls, screening for adversarial behavior.

The decision-model category Google is joining has gotten crowded fast. OpenAI shipped its own Decisions API for the Luna model in late September; TypeSafe's Jev, Cloudflare's open-source Clef, and AWS's Strands Decider 2B all target the same idea: cheap, calibrated, sub-second judgment for classification, routing and agent guardrails. Google's entry at $0.10 per million input tokens pushes the price floor for structured judgment lower still, and its benchmark-against-Jev choice signals the company sees the startup as the competitor to beat, not the general-purpose labs.

The economics argument is straightforward: an agent makes dozens of small judgments per task — which tool to call, whether a result is safe, whether to escalate to a human — and running each through a full generative model pays a latency and cost tax for text nobody needs. Offloading those to a calibrated classifier changes agent unit economics. What remains unproven is whether calibration holds on distributions outside Google's benchmark set, and whether Jev's lead on tool-calling judgments proves more important in production than Google's pricing.