Google Officially Launches Gemini Decision 1 and the Decisions API: Up to 1,000 Questions per Pass in 150 to 250 ms, Leads 7 of 9 Evals With the Lowest Calibration Error, $0.10 per Million Input Tokens and No Output Charges

Background and the pain point

Agents and automation are full of closed questions: which model should take a request, whether a tool call is safe, whether a parcel in a photo is damaged. Having a generative model write text and then parsing it is slow, brittle and gives no trustworthy probability. OpenAI and others shipped decision endpoints this autumn. Google's model first surfaced in the Gemini API changelog on November 11 (our first report was based on that). Hours later Sundar Pichai, @Google and Logan Kilpatrick made it official, so the claim that it had been pulled does not hold.

How it works

Google says Gemini Decision 1 is built on Gemini's native multimodal architecture and scores every candidate in a single forward pass without token-by-token decoding. There are three question types: noul (yes/no), choice (one of 2 to 255 options) and score (ordered rubric). Every answer carries a 0 to 1 confidence. Questions are independent, and the shared input is billed once per request.

Benchmarks

All figures are Google's. It ranks first on 7 of 9 evals: JevBench 92.2% (GPT-6 Luna 90.9%), Amazon ESCI 58.7%, PubMedQA 79.8%, Global-MMLU-Lite 86.0% and Video-MME 69.1%. It trails TypeSafe Jev 1.14 on Mind2Web Multimodal (86.2% vs 86.8%) and When2Call (79.9% vs 82.0%). Its Top-1 calibration error across 7 shared evals is 0.0250, the lowest of the four, and its total cost is $0.88 vs $4.14 for Flash-Lite, though TypeSafe costs only $0.68. Artificial Analysis has since published independent results: 66 on its Decision Index v2.1, level with Elephant 9B and Jev 1.14 and 7 points behind leaders Drex Max v1.1 and GLiDE v1.2 (both 73). It leads knowledge and reasoning evals but gets all answers right on only 33% of documents in AA-Paperclip vs 63% for Drex Max, and costs $345 per million tasks, about 3x Drex Max ($103).

Getting started

Get an API key in AI Studio and call client.decisions.create on v1beta in google-genai with model gemini-decision-1. You can start with Google's Doorstep Parcels and Speed Sketch demo apps. Google says Gemini Notebook uses it for model routing and cut everyday latency by up to 40%. OpenRouter, LangChain, LiteLLM, Pipecat, LiveKit and OpenCode integrations are rolling out. Set thresholds on the confidence score and send uncertain requests to a human.