A lot of what AI does inside apps is not writing at all. It is choosing: which model should handle this query, does this image break a rule, how severe is this complaint. Google's new Gemini Decisions API is built for that kind of work. Instead of generating an answer word by word, its model, Gemini Decision 1, scores a fixed set of options in one pass and returns probabilities, which makes it faster and much cheaper than asking a chat model.
The mechanics are simple to describe. You send one input, which can mix text, images, audio, video and PDFs up to 256K tokens, along with up to 1,000 independent questions. Each question is one of three types: a yes or no check, a choice among 2 to 255 options, or a score on an ordered rubric. Every answer comes with a confidence score between 0 and 1, so developers can send uncertain cases to a person or a stronger model. Google quotes 150-250 ms end to end, staying nearly flat from 1 to 100 questions per request on inputs up to 16K tokens.
Pricing is the headline. Decision 1 costs $0.10 per million input tokens and nothing for output, the shared input is billed once however many questions are asked, and there is a no-cost tier in AI Studio for prototyping. Google says that makes it nearly five times cheaper across its benchmark suite than using Gemini 3.5 Flash-Lite through the ordinary API, and claims the lowest calibration error on its benchmarks. Inside Google it already routes queries in Gemini Notebook, which Google says cut latency by up to 40% for everyday questions, and picks emoji reactions from more than 200 options in Google Labs' CC.
This is a category we already track. Jev, from TypeSafe AI, which we review, offers the same idea of typed choices and probabilities at $0.042 per million input tokens with free output, less than half Google's price. Google's advantages are native multimodal input and integrations across OpenRouter, LangChain, Vercel, LiteLLM and others; Jev's is cost. For teams running classification at scale, the comparison is worth doing on their own data.
The caveats are the usual ones. The benchmark results and calibration claims are Google's own, and the suite is not described in detail in the announcement. A decision model only chooses among the options you give it, so a badly framed question still gets a confident answer. And the latency figures assume inputs far shorter than the 256K maximum.