The takeaway
Google is selling the small judgments inside agent pipelines as a cheap, calibrated API primitive rather than another chat model.
Why it matters for builders
Replace router, tool-selection and grading prompts with a scoring call: the criteria become data, the confidence score gives you a cutoff for falling back to a stronger model, and output tokens stop being billed.
Gemini Decisions API Turns Agent Routing Into a Cheap Call
Google shipped a model on November 11 that never writes an answer. Gemini Decision 1 scores options instead, and the Decisions API wraps it as a primitive for the boring decisions that agent stacks currently burn full language model calls on: which tool to call, which model tier to route to, whether a policy check passes, how good a retrieved chunk is.
What the Decisions API actually does
You send one multimodal input - text, images, audio, video or PDFs up to 256K tokens - plus up to 1,000 independent questions. Instead of decoding tokens, the model evaluates every candidate option in a single forward pass and returns probabilities with a confidence score between 0 and 1.
Three question shapes are supported: yes/no probability estimates, multiple choice with 2 to 255 options returning normalized probabilities for every candidate, and ordered rubric scoring that returns a continuous weighted score with per-level probabilities. Google reports 150 to 250 milliseconds end to end, with median model compute of 50 to 100 milliseconds.
Pricing is the part builders will notice: $0.10 per 1M input tokens and nothing for output tokens, because there are none. The shared input is billed once per request no matter how many questions ride along. Google pegs it at roughly five times cheaper than Gemini 3.5 Flash-Lite through the Interactions API on its shared benchmark suite, and says Gemini Decision 1 leads that suite with the lowest calibration error. Those are Google's own numbers, not independently verified.
Already running inside Google
Google says the API is live in its own products as a low-latency control plane. Gemini Notebook uses it as a real-time router that classifies intent in milliseconds and sends everyday queries to cheaper tiers, claiming up to 40% lower latency. Stitch's rapid canvas experiment and Google Labs' CC emoji picker run on it, and Search uses the Decisions API as a safety layer that detects adversarial behavior in MCP tools.

Why the plumbing matters more than the model
Agent pipelines are mostly small judgments, and until now each one cost a full generation: a prompt, a schema, a parse, a retry when the model wrapped the answer in prose. That is where latency and token bills accumulate. A dedicated scoring endpoint with calibrated confidence turns those judgments into something you can budget, cache and fail over on.
The confidence score is the practically useful half. Your code can accept answers above a threshold and route ambiguous cases to a stronger model or a human, without a second prompt asking the same model whether it was sure.
Ecosystem support is rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode. For teams already paying per-token for routing logic inside n8n agents or custom orchestrators, the comparison is a straight swap: replace the router prompt with a scoring call, keep the criteria as data, and check whether the confidence threshold holds on your own traffic.
The Automation Brief
Read 5 AI stories instead of 50.
The essential moves in AI agents, models, automation and infrastructure — filtered for builders and operators, with the part that actually matters.
No noise. Unsubscribe anytime.
Editorial notes
Stefan Trbojevic
n8n Lab Editorial
12 November 2026
12 November 2026
AI disclosure: AI assisted with research and drafting. Factual claims are reviewed by an editor.



