Google launches Gemini Decisions API for one-pass scoring of up to 1,000 closed-set choices

Google ships a model built for choosing, not writing

Google has launched the Gemini Decisions API, a new endpoint in the Gemini API and Google AI Studio built around a purpose-built model, Gemini Decision 1 (gemini-decision-1), according to dev.to. Rather than composing a reply one token at a time, the model evaluates a fixed set of candidate options in one forward pass and hands back probabilities together with a confidence score. The announcement is dated November 11, 2026.

One shared input, up to 1,000 questions

A single request takes one shared input that can mix text, images, audio, video and PDFs up to 256K tokens, and as many as 1,000 separate questions can be answered against that input in the same call.

dev.to reports three question types: a binary yes/no probability; multiple choice with two to 255 options, where every candidate receives a normalized probability; and ordered rubric scoring, which returns a continuous weighted score plus probabilities for each rubric level. Every answer also carries a confidence score between 0 and 1, giving application code a signal for routing uncertain cases to a fallback or a human reviewer.

The use cases Google names are tool selection, policy checks, model routing and context grading — all decision points where the answer comes from a known set rather than free-form prose.

Latency and pricing

Per Google's figures as relayed by dev.to, end-to-end latency runs 150–250 ms, with median model compute of 50–100 ms on contexts up to 16K tokens, and latency stays nearly flat from 1 to 100 questions per request.

Pricing is USD 0.10 per million input tokens, and nothing is billed for output. Google puts this at nearly five times cheaper than Gemini 3.5 Flash-Lite used with the Interactions API across a shared benchmark suite. A no-cost tier is available in Google AI Studio. Google also claims the model leads its benchmarks in routing, reasoning, ranking and multimodal understanding, with the lowest calibration error among the models compared.

Already running inside Google

Google says the model is in production internally: it routes requests between models in Gemini Notebook, cutting latency by up to 40 percent on daily queries; drives wireframing in Stitch; picks emojis in CC from a pool of more than 200 candidates; and acts as a safety layer for MCP tools in Google Search.

Third-party support is arriving through OpenRouter, LangChain, the Vercel AI SDK, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, alongside Google's google-genai SDKs.

Why it matters

Agent frameworks currently burn generation tokens to answer questions that are really classifications — which tool to call, whether a policy permits an action, which model should handle a prompt. A model that scores a closed set of options in one pass, returns calibrated probabilities and costs a tenth of a dollar per million input tokens changes the arithmetic on those calls. The built-in confidence score also gives developers a defensible threshold for escalation instead of string-matching a model's reply. The breadth of integrations announced at launch suggests the agent-tooling ecosystem is treating this as infrastructure rather than a curiosity.