Google’s Gemini Decisions API picks answers instead of writing them, for $0.10 per million tokens

Google launched the Gemini Decisions API, powered by a new model called Gemini Decision 1. Instead of generating text, it scores a fixed set of options in one pass and returns probabilities and confidence scores. Google says it takes 150 to 250 milliseconds end to end, costs $0.10 per million input tokens, and bills nothing for output. It handles text, images, audio, video and PDFs, and one request can carry up to 1,000 yes/no, multiple-choice or rubric-scoring questions.

It is available in Google AI Studio with a no-cost prototyping tier, with integrations rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI and others. Google says it is roughly five times cheaper on its benchmark suite than Gemini 3.5 Flash-Lite via the Interactions API, and it already uses the model for routing in Gemini Notebook and as a safety layer for tools in Search.

Much of what agents do isn’t creative writing. It is small, bounded choices: which tool to call, whether a request breaks a policy, which model should handle a query. Using a full generative model for those is slow and pricey. A purpose-built picker that returns a number with a confidence score is the right tool. The idea isn’t unique to Google: TypeSafe’s Jev pioneered typed decisions with calibrated confidence (coding guide), OpenAI reportedly has its own Decisions API for policing agents (coverage), and Meta’s Muse Ember targets the same cheap agent-call market (coverage). Google’s newest efficient generalist, Gemini 4 Hydrogen, is the execution side of the same multi-model strategy.

Google pitches the confidence score as a way to send ambiguous cases to a fallback or a human reviewer. That is a sane pattern: systems that know when they’re unsure are safer than systems that always sound sure. Google claims the lowest calibration error on its benchmark suite, but those are its own numbers and outside testing is needed.

The launch example is a breakroom spill, but the same API could score insurance claims, call recordings or content moderation, areas where a confident-looking probability can hide bias or error. Decision models could make moderation more flexible and transparent, or make it far easier to label everything (analysis). If teams treat a score as an answer and skip the human route for borderline cases, the speed becomes a liability.

The likelier path is that this becomes invisible plumbing inside many agent stacks. That makes the quality of defaults, and how builders set thresholds, matter more than headline latency. More in our Google Gemini coverage, agents coverage and tools coverage.