{
  "id": 19885929,
  "title": "Benchmarking the Gemini Decisions API",
  "url": "https://urgent.news/2026/11/13/benchmarking-the-gemini-decisions-api",
  "topic": "ai",
  "section": "AI",
  "published": "2026-11-13T23:45:04.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/gde/benchmarking-the-gemini-decisions-api-l31"
  },
  "original_language": "en",
  "account": "This article presents a detailed walkthrough of the Gemini Decisions API, utilizing it to triage customer support tickets. An agent built on the API is tested against four generative models - Gemini 3.8 Flash, Gemini 4 Hydrogen, Gemini 3.5 Flash-Lite, and Gemma 4 E2B on a laptop CPU. All five models are compared based on accuracy, speed, and cost.\n\nThe Gemini Decisions API is designed to handle fast, closed-set decisions such as routing, policy checks, model selection, and grading. It accepts one input (text, images, audio, video or PDF) and up to 1,000 questions, returning probabilities for each allowed answer. The model gemini-decision-1 scores every allowed answer in one forward pass and outputs probabilities. It does not generate text, so the cost is minimal at $0.10 per million input tokens.\n\nThe API supports three types of questions: yes or no (noul), pick one of multiple options (choice) with probabilities per option and a confidence score, and a level on an ordered rubric (score). The confidence score indicates the probability that the top answer is correct.\n\nThe demo uses customer-support tickets as examples, with each ticket being asked three questions: which team owns it, if the customer is asking for a refund, and how urgent the issue is. These questions are sent to five backends through one function in decide.py, which scores and times the responses.\n\nThe models tested are Gemini 3.8 Flash, Gemini 4 Hydrogen, Gemini 3.5 Flash-Lite, and Gemma 4 E2B on a laptop CPU. The four generative models are queried using REST, with the same inputs and answer shapes. The Decisions API, gemini-decision-1, is called over REST, and the results are compared across the four models.\n\nThe article also provides a step-by-step guide on how to make one call with curl and another method to call the API from Python over REST. Additionally, it explains how to run the three test scripts: make bench, make scale, and make web.",
  "summary": "This article provides a step by step guide to the new Gemini Decisions API, using customer-support ticket triage as the example. An ADK agent is built on it, and the same questions are put to four generative models, Gemini 3.8 Flash, Gemini 4 Hydrogen, Gemini 3.5 Flash-Lite and Gemma 4 E2B on a laptop CPU, so all five can be compared on accuracy, speed and cost.…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}