Google has launched a Gemini Decisions API, powered by a new model, Gemini Decision 1, MarkTechPost reports. It isn’t a chatbot. You send one shared input plus up to 1,000 typed questions, and you get back calibrated probabilities, not paragraphs. Google is pitching it for the little branching calls inside agents: routing, policy checks, tool selection and grading. Pricing is $0.10 per million input tokens, with no output tokens billed.
The shared context is 262,144 tokens, and the model can read text, images, video, audio and PDFs in one prefill pass. There are three question types: noul (yes/no, with a probability bounded between 0.01 and 0.99), choice (one of 2 to 255 options, with a probability for each), and score (a weighted score across a rubric). Every answer includes a confidence value from 0 to 1, and Google suggests sending low-confidence cases to a fallback or a person. Questions are scored in parallel and can’t see each other.
It’s Google-hosted only, on the Gemini Developer API (v1beta). There are no open weights. Google’s own table has the model leading 7 of 9 accuracy benchmarks, with its best reported score 92.2% on JevBench and a weaker 58.7% Macro-F1 on Amazon ESCI. MarkTechPost also notes it isn’t the cheapest option in that comparison set.
This is the same general idea as OpenAI’s Decisions API, which I wrote about when it left beta, only on Gemini’s stack.
My take
I like this shape for agent plumbing. A lot of what agents do all day is “is this spam?” or “which tool should I call?”, and a calibrated number is easier to test than a paragraph that might wander. I’d still keep a hard rule or a human in front of anything that sends, buys or deletes. A fast decision can still be wrong.
God Bless, and May the Source Be with You!
