# Gemini Decisions API The Gemini Decisions API is [[Google]]'s [[Decision Models (DMs)|decision model]] endpoint, announced on 11 November 2026 by [[Google DeepMind]]. It runs a new model, Gemini Decision 1 (`gemini-decision-1`). You give it an input and a set of bounded questions, and instead of writing an answer token by token, it scores every option in a single forward pass and returns probabilities. Why should you care? Because it confirms that decision models are a real category, not a one-company trick. [[Jev]] opened it in September 2026, [[Microsoft-Decision-1]] followed in October, OpenAI has a Decisions API in beta, and now Google. As one Hacker News commenter put it: "Looks like Google announced their own Jev inspired API and model." Google's announcement doesn't name Jev, though. ## How it works - One shared input per request: text, images, audio, video and PDFs, up to 256K tokens - Up to 1,000 independent questions about that input in one request - Three question types: - **noul** (yes/no): the probability that a condition holds. The name follows Jev's naming ("Bernoulli") - **choice**: pick one of 2 to 255 options, with normalized probabilities for each - **score**: grade the input against an ordered rubric, with a weighted score and per-level probabilities - Every answer comes with a confidence score from 0 to 1, so your code can send ambiguous cases to a fallback or a human The multimodal part is what sets it apart. Commenters called it "the first decision AI that actually handles audio, video, and PDF". Grade a scanned PDF claim against damage photos, or check a call recording against account data, without paying for transcription or OCR first. ## Pricing and latency - $0.10 per million input tokens, $0 for output. No thinking tokens, no transcription or OCR fees - The shared input is billed once per request, whatever the number of questions; each question only adds its own instruction tokens - 150-250 ms end to end (50-100 ms of model compute), nearly flat from 1 to 100 questions per request, for inputs up to 16K tokens - Free tier in [[Google AI Studio]] for prototyping Google says it's nearly 5x cheaper than asking Gemini 3.5 Flash-Lite the same questions, and that it has "the lowest calibration error" on its benchmark suite. On Hacker News, people found it pricey compared with other decision models ($0.10 vs $0.02 per million tokens for one alternative). One commenter measured a Brier score about 4x lower than Jev (better calibrated) and faster p50 latency, but at twice Jev's end-to-end cost. ## Where Google uses it - Gemini Notebook: model routing, up to 40% lower latency on everyday queries - Google Search: a safety layer that detects adversarial behavior in MCP tools - Google Labs experiments: picking an emoji reaction among 200+ options, or a voice-driven wireframe canvas (Stitch) Integrations are rolling out for [[OpenRouter]], LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode. Google Cloud support is "coming soon"; nothing on [[Vertex AI]] yet. ## Reception on Hacker News The thread (568 points, 161 comments) was mostly about decision models in general: - **Why did it take so long?** Reading a probability out of a model is not new. Some think vendors had no incentive to sell something that cuts token use by orders of magnitude. Others say it's branding: Jev gave the idea a name - **Calibration is the whole game.** "LLM token probabilities are really poorly calibrated for decision making", and a model calibrated on the vendor's data isn't automatically calibrated on yours (see [[AI Model Calibration]]) - **No reasoning.** "no other tokens means I can't get reasoning or a justification" - **Opaque classifiers as a service.** One commenter warned about using it for mortgage decisions and similar cases, where a confident number hides discrimination - **Trust in Google.** People brought up products Google killed (Jules) and a prompt blocked by the prohibited-use policy ## My take I think this is one of the most useful things Google shipped this year. A lot of what agents do all day is deciding: which tool, is this relevant, is the agent stuck, does a human need to look at this. Doing that with a text generator is slow and wasteful. The audio, video and PDF support fills a real gap. The usual caveat applies: measure calibration on YOUR data before you trust any threshold. And split big judgments into many small questions ([[Atomic Question Decomposition]]), which is exactly what 1,000 questions per request is made for. ## References - Google, "Gemini Decisions API: Fast, calibrated multimodal decisions" (Min Choi and Felix Yu, 2026-11-11): https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/ - Documentation: https://aistudio.google.com/docs/decisions-api - Hacker News discussion: https://news.ycombinator.com/item?id=50456791 ## Related - [[Decision Models (DMs)]] - [[Jev]] - [[Microsoft-Decision-1]] - [[Laya]] - [[AI Model Calibration]] - [[Atomic Question Decomposition]] - [[Model routing]] - [[Gemini]] - [[Google AI Studio]]