Google Launches the Gemini Decisions API: A Model That Writes Nothing and Resolves Up to 1,000 Questions Per Call

Google is stepping into the race for “decision models.” On November 11, Google DeepMind unveiled the Gemini Decisions API, a new endpoint for the Gemini API powered by a dedicated model, Gemini Decision 1. Its core concept: write nothing, but decide.

Probabilities Instead of Text

A traditional generative model answers word by word. When a software program merely needs to know whether to invoke a specific tool or whether a message violates a policy, it still has to parse that text—adding latency and interpretation errors. Gemini Decision 1 takes the opposite approach: it takes an input alongside a list of closed-ended questions, evaluates all possible answers in a single pass, and returns a probability and a confidence score for each.

The documentation outlines three question types: yes/no, multiple-choice from 2 to 255 options, and a score along an ordered scale. A single request can bundle up to 1,000 questions over a shared input that mixes text, images, audio, video, and PDFs. Questions are evaluated independently without seeing one another; Google recommends asking narrow questions and combining them downstream in code using standard conditional logic.

What Google Promises—and What It Skips

The advertised figures are aimed squarely at agentic workloads: 150 to 250 milliseconds end-to-end, almost the same for one question as for a hundred, and a price tag of $0.10 per million input tokens, with zero output tokens billed. The shared input is counted only once, regardless of question volume. Google highlights internal use cases: routing queries to the right model in Gemini Notebook (up to 40% lower latency on common requests, according to the company), selecting an emoji from over 200 options in the CC app, and detecting malicious behavior for Google Search MCP tools.

The model card also makes clear what it cannot do: no text generation, no function calling, no Batch API, and no caching. It is a router, not an assistant.

Google claims its model tops its internal benchmark suite with the best calibration, while costing nearly five times less than Gemini 3.5 Flash-Lite across those same evaluations.

A More Nuanced First Independent Verdict

That same day, Artificial Analysis published its own measurements. On its Decision Index v2.1, which blends accuracy and calibration, Gemini Decision 1 scored 66 points: a solid result, yet behind Drex Max v1.1 and GLiDE v1.2 (73 each), as well as several other specialized decision models. However, it took the top spot among decision models on knowledge and reasoning tasks. On speed, it handled a task in an average of 211 milliseconds, right within Google’s stated range. On price, the comparison flips: $345 per million tasks on the index, versus $103 for Drex Max v1.1, $68 for Decider v1.2, and $14 for Celeris-1.1 Decision. Cheaper than a generative model, certainly, but far from the cheapest in its class.