LRLLM Results

Gemini Decisions API: What Google Confirmed

Google announced a fast closed-set decision endpoint on November 11, 2026. The price, limits and availability are stated by Google; the latency and benchmark results are its own claims.

Published November 12, 2026 · LLM Results. Everything below comes from Google's Gemini Decisions API announcement, dated November 11, 2026. We have not called the API ourselves. Statements are marked as Google's claims where they are not independently checkable.

What Google States

  • What it is: a Gemini API endpoint for fast, closed-set decisions such as tool selection, policy checks, model routing and context grading. It is powered by a model Google calls Gemini Decision 1 (gemini-decision-1).
  • Inputs: one multimodal input (text, images, audio, video or PDFs, up to 256K tokens) plus up to 1,000 questions per request.
  • Question types: yes/no probability checks, multiple choice with 2 to 255 options, and ordered rubric scoring. Each answer carries a confidence score from 0 to 1, so uncertain cases can go to a fallback or a human reviewer.
  • Method: options are scored in a single forward pass rather than by generating text token by token.
  • Price: $0.10 per million input tokens and no output-token charge. The shared input is billed once per request; each question adds only its own instruction and criteria tokens. A no-cost tier is offered in Google AI Studio.
  • Availability: the Gemini API and Google AI Studio, with google-genai SDK support. Google lists integrations rolling out with OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode. Gemini in Google Cloud is listed as coming soon.

Company-Reported Results

  • Google reports 150 to 250 ms end-to-end latency, with 50 to 100 ms median model compute, and says latency stays nearly flat from 1 to 100 questions per request for contexts up to 16K tokens.
  • Google says the model leads a benchmark suite covering agent routing, reasoning, ranking and multimodal understanding, with the lowest calibration error, and is nearly 5x cheaper across that suite than Gemini 3.5 Flash-Lite used with the Interactions API. These figures are Google's own and we have not reproduced them.
  • Google cites early internal uses: model routing in Gemini Notebook (up to 40% lower latency for everyday queries), wireframe generation in Stitch, emoji reaction selection in a Google Labs product, and a safety layer for MCP tools in Google Search. These are company descriptions of its own products.

What It Means for AI Users

Our inference: this is a developer building block rather than a chat model. It suits the cheap, repeated decisions inside agent and retrieval pipelines, such as routing a query or grading whether retrieved context is relevant, where a full generative call is slow and costly. The confidence score is the useful design feature, because it makes escalation rules explicit. We have not tested calibration or accuracy on any task, so evaluate against labelled examples before relying on it, and compare against a small generative model on the same job.

For AI search visibility the direct effect is limited: Google does not say this changes how Google Search ranks or cites pages. It says Search uses the model as a safety layer for MCP tools, which describes tool-use safety rather than result ranking. We would not draw GEO conclusions from this release.