Google on Tuesday introduced the Gemini Decisions API, a new endpoint built for fast, closed-set decisions, powered by a purpose-built model called Gemini Decision 1 (gemini-decision-1). The API is available now through the Gemini API in Google AI Studio, according to Google's announcement.
The pitch is simple: a large class of production AI work does not need a model that writes essays. Tool selection, policy checks, model routing and context grading are all bounded choices — and for those, generative text generation is slow, expensive and poorly calibrated. The Decisions API instead takes your candidate options and scores them, returning probabilities and confidence scores rather than generated answer text.
For more context on developer tools like this one, follow our AI tools coverage.
One input, up to 1,000 questions, one forward pass
The API's input model is unusual and deliberately so. Developers pass one multimodal input — text, images, audio, video or PDFs, with contexts up to 256K tokens — along with up to 1,000 independent questions in a single request. Questions can mix yes/no checks, multiple-choice selections and rubric scores against the same shared input.
Because Gemini Decision 1 scores the candidate options in a single forward pass rather than decoding tokens one by one, Google reports median model compute of 50–100 milliseconds and end-to-end latency of 150–250 milliseconds. The company says performance stays nearly flat as questions scale from 1 to 100 per request, and that the shared input is billed once no matter how many questions evaluate against it.
Pricing: no output tokens, ever
Gemini Decision 1 is priced at $0.10 per 1 million input tokens, with zero output tokens billed — a meaningful structural difference from generative APIs, where output tokens usually dominate the bill. Google says that makes the Decisions API nearly 5x cheaper across its shared benchmark suite than Gemini 3.5 Flash-Lite used through the Interactions API, with no separate audio transcription or OCR charges. A no-cost tier is available in Google AI Studio for prototyping.
Calibration is the headline claim
Google's announcement leads with a claim about calibration: across benchmarks spanning agent routing, reasoning, ranking and multimodal understanding, the company says Gemini Decision 1 leads the suite with the lowest calibration error. Every answer ships with a confidence score from 0 to 1, so application code can route ambiguous cases to a fallback model or a human reviewer instead of trusting a low-confidence classification.
Calibration matters because routing failures compound in agent systems. An agent that picks the wrong tool with high confidence fails silently; an agent that knows when it is unsure can degrade gracefully. Making calibrated confidence a first-class API output — alongside the classification itself — is arguably the most production-minded part of the launch.
Native multimodality in a single pass
Built on Gemini's native multimodal architecture, the model processes text, images, audio, video and PDFs directly in one forward pass. Google's examples show the mixing of modalities in practice: pairing a call recording with account metadata, or a scanned PDF insurance claim with damage photos, and tuning resolution on image and video inputs to balance visual detail against token cost.
The company also highlighted internal adoption. Google says teams across the company already use the Decisions API as a low-latency control plane, shipping features by defining options rather than training custom classifiers — including Stitch Rapid, which the announcement says uses the API to design an entire product flow in seconds.
Ecosystem support from day one
Google listed launch-day integrations across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, alongside the google-genai SDKs for Python, JavaScript and REST. Support in Gemini on Google Cloud is coming soon, according to the announcement.
Developers can start today with an API key in Google AI Studio, and Google published demo apps — Doorstep Parcels, Speed Sketch and Find the Zoo Keeper — to show the request patterns in action.
Why it matters for the agent economy
The Decisions API targets the least glamorous but most heavily exercised layer of the emerging agent stack: the constant small decisions an agent makes dozens of times per task. If routing, policy checks and context grading can run at 150–250 milliseconds with predictable pricing, the economics of long-running agent workflows change materially — the decision layer stops being a generative-model bottleneck and becomes closer to traditional infrastructure.
The launch also sharpens Google's positioning against rivals building dedicated classification and routing models, by attaching frontier-lab multimodal capability to a commodity price point. It also signals where Google sees the API market splitting: generative models for open-ended work, decision models for the bounded, high-volume checks that keep agent systems on rails. Whether independent benchmarks confirm the calibration claims will be the metric to watch as adoption grows.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
