Google Launches Gemini Decision 1, a Model That Answers Instead of Writing

Google launched the Gemini Decisions API on November 11, 2026, powered by a new model called Gemini Decision 1. Unlike the chat models that dominate AI headlines, this one does not write. It answers bounded questions — yes or no, multiple choice, or a score against a rubric — by scoring the candidate options in a single forward pass and returning calibrated probabilities and confidence scores. Google reports end-to-end latency of 150 to 250 milliseconds, and pricing of $0.10 per million input tokens with zero output tokens billed.

What Happened

Google announced the new endpoint in a post on its developer blog on November 11, 2026, written by Min Choi, a product manager at Google DeepMind, and Felix Yu, a principal scientist there. The Decisions API is available through the Gemini API in Google AI Studio, on the Gemini Developer API v1beta. A no-cost tier is available in Google AI Studio for prototyping; Vertex AI support is coming soon.

A single request takes one multimodal input of up to 256,000 tokens — text, images, audio, video and PDFs — together with up to 1,000 independent questions. Gemini Decision 1 scores the candidate options in one forward pass instead of generating answer text token by token. Google says the model keeps median latency at 50 to 100 milliseconds of model compute, or 150 to 250 milliseconds end to end, across multimodal inputs and contexts up to 16,000 tokens, remaining nearly flat from 1 to 100 questions per request.

The endpoint is built for bounded choices such as tool selection, policy checks, model routing, context grading and ranking. Google says it led its own benchmark suite with the lowest calibration error and is nearly five times cheaper than Gemini 3.5 Flash-Lite with the Interactions API; those are Google's own figures.

What a Decision Model Is

Most AI models in everyday use are generative: they produce text, images or code one token at a time. That is powerful but wasteful when the real question is a choice — is this document relevant, does this image violate policy, which tool should handle this request? A generative model answers such questions by writing out reasoning and an answer, burning time and tokens on prose nobody reads.

A decision model skips the prose. It takes the input and the list of possible answers, and returns a probability for each one. Because the model makes a single forward pass instead of generating tokens, it is fast and cheap — and because its outputs are calibrated probabilities rather than confident-sounding text, they can be used in automated systems with explicit risk thresholds.

Why It Matters

The launch is a sign of how AI infrastructure is maturing. The first wave of the AI boom was about models that could write anything; the current wave is about making those models cheap, fast and reliable enough to sit inside real software. Every agent system, every retrieval pipeline and every content-moderation flow needs thousands of small judgements a day — which model should answer, whether a tool call is safe, whether a document is relevant. Doing each judgement with a full generative model is slow and expensive. A purpose-built decision model turns that into a commodity operation.

The economics are striking. At $0.10 per million input tokens with no output tokens billed, and with the shared input billed once for up to 1,000 questions, a single request can carry a whole battery of checks for the price of reading the input once.

Limits to note: the endpoint runs on the Developer API v1beta only, with no Vertex AI listing yet, and independent benchmarking by Artificial Analysis ranked it ninth of eighteen decision models — competitive, not dominant.

What This Means for Founders and Investors

For founders building agents and AI workflows, Gemini Decision 1 is a practical tool to evaluate now. Routing, guardrails, evaluation and retrieval grading are exactly the jobs that eat inference budgets, and a fast, cheap decision endpoint with a free tier for prototyping lowers the cost of building reliable agents.

For investors, the launch illustrates where value is moving in the AI stack. The frontier labs are no longer competing only on the biggest model; they are competing on the full shelf of specialised models — fast, cheap, calibrated components that enterprises can compose into systems. The companies that win enterprise AI budgets will be the ones that offer the whole shelf, with pricing and latency that make automation cheaper than the human alternative.

Angeled.com connects founders with investors.