Google's Gemini Decisions API Creates a New AI Service Business: Replace Expensive Full-Generation Calls With Cheap, Near-Instant Bounded Decisions.

Most teams building with LLMs are overpaying for the wrong kind of answer.

Google's new Gemini Decisions API, announced November 11, 2026, names the exact waste.

The pitch from Google DeepMind's Min Choi and Felix Yu is blunt: the API can "evaluate up to 1,000 yes/no, choice, and rubric-scoring questions in a fast, single pass across text, images, audio, video, and PDFs."

What the API actually does differently

Gemini Decision 1, the model behind the API, "scores your candidate options in a single forward pass, returning probabilities and confidence scores in 150-250 ms end-to-end latency." There's no output token generation at all — the model picks among options you give it and returns a confidence score, not prose.

That architecture choice changes the economics directly. Google states: "With zero output tokens billed, Gemini Decision 1 with the Decisions API is nearly 5x cheaper across the shared benchmark suite than Gemini 3.5 Flash-Lite."

Pricing is "$0.10 per 1M input tokens and $0 for output tokens," with a no-cost tier in Google AI Studio for prototyping. Google's own teams are already using it for things like dynamic model routing, emoji selection, and safety detection in Search — all classic bounded-choice problems that most companies currently solve with full generative calls.

Why this is a service, not just a model announcement

Nobody ships an AI pipeline with a line item labeled "bounded choice masquerading as generation." It accumulates quietly: a moderation check that calls a chat model and parses the first word of the reply, a routing step that generates a sentence and regexes out a category, an eval harness that asks an LLM to "respond with only PASS or FAIL" and still pays for a full completion.

Each of those is a candidate for migration the moment a team sees the cost and latency gap named out loud, with a vendor-published number behind it.

The pitch that sells itself

Open with the gap, not a generic "optimize your AI costs" line: "Google just shipped an API built specifically for the yes/no and routing calls buried in your pipeline — 150-250ms, nearly 5x cheaper than generation. We'll find every call in your stack that's secretly a bounded choice and migrate just those, nothing else."

That's concrete, cites Google's own numbers, and gives the client an audit they can verify by rerunning the comparison themselves.

Bottom line

This works as a playbook because the waste isn't hypothetical — Google just built a product that only makes sense if lots of teams are currently overpaying for bounded decisions. Finding those calls, migrating only those, and leaving the rest of the pipeline alone is a narrow, fast, and easily-priced service.

Source: https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/