Google's Gemini Decisions API answers routing questions in one pass, no text generation

A model endpoint built for branching, not prose

Chat model APIs are designed to produce text that a person will read. When an application only needs to pick a tool, classify a request, or decide where to route it, the common workaround is to prompt a chat model for JSON and parse whatever comes back. Google says it has now built something for exactly that case: the Gemini Decisions API, announced in a post on dev.to, is a dedicated endpoint for control-flow decisions.

The endpoint runs on a new model called Gemini Decision 1. According to the post, it answers narrowly scoped questions in a single forward pass, generates no text tokens, and returns calibrated probabilities in around 200 milliseconds.

One shared context, up to 1,000 questions

A request pairs one shared input, which can mix text, images, audio, video, and PDFs, with as many as 1,000 independent questions. The questions are blind to one another and evaluated in parallel, so response time barely changes between a one-question call and a hundred-question call, the post claims.

Three question types can be combined in a single request: yes-or-no checks, multiple-choice selections, and ordered rubric-style ratings. Every answer carries a confidence score, which Google positions as the hook for routing logic: results the model is sure about can drive code paths directly, while uncertain ones can be escalated to a heavier reasoning model or a human.

Google says the endpoint already runs in production inside its own products. Gemini Notebook uses it to steer prompts to different model tiers, and Stitch uses it to turn spoken edit commands into live changes on wireframes. Those usage claims come from the announcement post and have not been independently verified.

Pricing and availability

Input is priced at $0.10 per million tokens, and the shared context is billed only once regardless of how many questions are attached to it. Output tokens are free, which fits a model that generates none. A free API key is available through Google AI Studio, where the documentation is also hosted.

Why it matters

The prompt-for-JSON pattern that most agents rely on today carries hidden costs. Every routing decision waits on autoregressive text generation, depends on parsing code that breaks when a model wraps its answer in markdown or adds a stray field, and spends output tokens on scaffolding nobody reads. A purpose-built decision endpoint removes all three.

If the speed and calibration claims hold up, the effect is architectural. Agent frameworks gain a cheap, fast gate to place in front of expensive reasoning models, escalating only when confidence drops. Application developers get multimodal classification over images, audio, and documents without training or hosting a separate model per task. Confidence-based escalation also becomes an API-level concept rather than something each team rebuilds from logprobs and guesswork.

The caveat is that everything here rests on a single announcement post. The flat-latency behaviour, the calibration quality, and the internal production usage are all Google's own statements, and developers will want to benchmark the endpoint against their existing prompt-based routing before rearchitecting around it.