Google launches Gemini Decision 1 for fast, cheap AI decisions

Google has launched the Gemini Decisions API, a new endpoint built around a dedicated model called Gemini Decision 1, for making fast, closed-set decisions. Developers can use it for bounded decisions such as picking the right tool, running policy checks, choosing which model to route a request to, and grading retrieved context. They can feed it one multimodal input — combining text, images, audio, video, and PDFs, up to 256,000 tokens — alongside as many as 1,000 independent questions in a single request. Instead of generating an answer word by word, Gemini Decision 1 scores the candidate options in one forward pass and returns probabilities and confidence scores in 150 to 250 milliseconds end to end, at $0.10 per million input tokens with no charge for output tokens.

Google says Gemini Decision 1 leads on benchmarks covering agent routing, reasoning, ranking, and multimodal understanding, with the lowest calibration error in the group, and that its latency stays close to flat — 50 to 100 milliseconds of model compute, across contexts up to 16,000 tokens — whether a request carries one question or a hundred. Because it bills no output tokens, Google says Gemini Decision 1 through the Decisions API works out to nearly five times cheaper than Gemini 3.5 Flash-Lite through the Interactions API on the same benchmark suite.

The model runs on Gemini’s native multimodal architecture, evaluating text, images, audio, video, and PDFs together in one pass, and can return up to three answer types in the same request: yes/no probability estimates, multiple-choice selection among 2 to 255 options, and ordered-rubric scores starting at zero — each with its own confidence score. Google highlights several of its own products already built on the API: Gemini Notebook uses it to route queries to the right model tier, which Google says cuts latency by up to 40% on everyday queries; Stitch uses it to generate and edit interactive product-flow wireframes in seconds; the Labs app CC uses it to pick the right reaction from a palette of more than 200 emoji almost instantly; and Search uses it as a safety layer to flag adversarial behavior against MCP tools.

Google says integrations are already rolling out across OpenRouter, Vercel, and LangChain, along with LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode, alongside its own google-genai SDKs, with support for Gemini on Google Cloud coming later.

Why it matters

A model built purely to score bounded options, not generate free text, could let AI products add fast decision-making — tool choice, routing, content classification — without the latency or output-token cost of a full generative call, if Google’s own benchmark comparisons hold up under independent testing.