Skip to content
RSS feed
thesmartaistack.
Blog

Blog

Blog

Gemini Decision 1: what Google's decision model is for, and when a regular Gemini model is the better call

Gemini Decision 1 answers up to 1,000 typed questions in one pass, with calibrated probabilities and zero text output. When to use it, when not.

By the editorial desk · Updated Nov 13, 2026

Google title card reading Gemini Decisions API in white type on a dark navy and cobalt gradient
Artwork: Google Google Blog announcement · Google's launch title card for the Decisions API, powered by Gemini Decision 1.

Google's newest Gemini model does not write a single word of text. Gemini Decision 1, announced alongside the Decisions API, answers up to 1,000 yes/no, multiple-choice, and rubric-scoring questions against one shared input in a single pass, returning calibrated probabilities and a confidence score instead of generated prose. Every figure here was read off Google's own announcement and the Gemini API documentation on November 13, 2026, and the model's job description, not its benchmark numbers, is what decides whether your project needs it.

What Gemini Decision 1 actually is, by Google's own description

Google's model page defines it in one sentence: Gemini Decision 1 is "a specialized multimodal model built for the Decisions API" that "evaluates typed questions over text, images, video, audio, and PDFs in a single prefill pass and returns calibrated outcome probabilities with zero output tokens." One request accepts a combined budget of up to 262,144 tokens covering the shared input and every question's instructions, and up to 1,000 independent questions within that budget. The three question types cover the shapes code actually branches on: yes/no checks return a calibrated probability from 0.01 to 0.99, single-select evaluates 2 to 255 options and returns the top pick with per-option probabilities, and rubric rating grades input across 2 to 255 ordered levels and returns a probability-weighted expected score.

The part that separates it from every other Gemini: there is no text generation at all. Output token usage is always zero. Where a regular model writes "yes, the package appears damaged" and your code parses that sentence, Decision 1 returns a number your code can branch on directly.

Google title card reading Gemini Decisions API in white type on a dark navy and cobalt gradient
Artwork: Google · Google Blog announcement · Google's launch title card for the Decisions API, powered by Gemini Decision 1.

Who this helps, and what it lets them do

It helps developers whose code needs a judgment to branch on, and Google's own teams are the first proof. The announcement names four internal deployments: Gemini Notebook uses it as a real-time router that "classifies intent in milliseconds to serve everyday queries with up to 40% lower latency"; Stitch uses it for rapid design iteration; CC uses it to pick from a palette of more than 200 emoji reactions; and Google Search applies it as a safety layer for MCP tools. For your own work, the model page lists the fitting jobs: request routing, policy checks across many rules at once, candidate scoring against a rubric, tool selection, and low-latency classification over live media streams.

The economics reward batching questions into one call. Shared input is billed once per request no matter how many questions ride on it, each question adds only its own instruction and criteria tokens, and output is always $0. At $0.10 per 1M input tokens on the paid tier, the shared document is billed once, but each of the 50 questions still adds its own instruction and criteria tokens to the bill. A free tier covers prototyping in Google AI Studio.

When it earns its place in a build

Three signals say Decision 1 is the right tool. First, the decision is repetitive and typed: the same yes/no check, the same choice among known options, or the same rubric, applied over and over to different inputs. Google's own figure of 150 to 250 ms end-to-end latency, measured on multimodal inputs at contexts up to 16K tokens, at $0.10 per 1M input tokens with zero output billed, is what a control plane needs; a generative model that writes a sentence per decision pays for tokens nobody wanted. Second, the input is already multimodal: a call recording next to account metadata, or a scanned PDF claim next to damage photos, evaluated in one pass. Third, the decision volume is high enough that calibration matters: every answer carries a confidence score from 0 to 1, so code can route ambiguous cases to a fallback or human reviewer instead of acting on a guess. If you automate routing on that confidence, set the threshold from labeled examples of your own application and escalate the cases below it; the docs' confidence section is calibration guidance, not a universal cutoff.

Rate limits back the control-plane framing. The free tier allows 200 requests per minute and 10,000 per day; Tier 1 raises that to 1,500 RPM and 50,000 RPD; Tier 3 reaches 15,000 RPM with unlimited daily requests. Google reports median model compute of 50 to 100 ms, staying nearly flat from 1 to 100 questions per request; how latency behaves past 100 questions is not a figure Google publishes.

When a regular Gemini model is already enough

The trade-off is scope, and it is sharper than most model comparisons because the capabilities simply are not there. Decision 1 does not generate text, so it cannot answer an open question, draft anything, or reason at length. The model page marks Batch API, context caching, code execution, computer use, file search, Flex inference, function calling, and audio generation as not supported. If your job is "have a conversation" or "produce content", Carbon or Hydrogen is the actual product; this model cannot play that role at any price.

Two more cases keep a general model in place. If the decision changes often and the criteria live in your head rather than in written rubric levels, maintaining typed questions adds ceremony a prompt to a regular model does not need. And low-volume pipelines that already call a general model for the surrounding task can often fold a small judgment into the same call rather than paying a second request's round trip.

How it relates to Carbon and Hydrogen, and where the claims stand

Google does not position Decision 1 as a sibling in the reasoning lineup, and the benchmark comparison it does publish is specific: nearly 5x cheaper across a shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API, with what Google calls the lowest calibration error on benchmarks spanning agent routing, reasoning, ranking, and multimodal understanding. Both are Google's own reported numbers, and no public third-party leaderboard entry for gemini-decision-1 was found at check time, so this article treats them as vendor claims. What the pricing table does establish on its own: $0.10 per 1M input tokens, $0 output, free-tier prototyping, and no Batch API support.

Every figure in this piece is Google's published number with the date we checked it. The claims worth re-reading as the model matures: the latency and cost figures above, the benchmark lead, and the internal deployment list, all from the announcement and documentation pages dated November 13, 2026.

Getting started

Grab an API key in Google AI Studio, try the Decisions Playground, Doorstep Parcels, or Find the Zookeeper sample apps, and follow the Python, JavaScript, and REST guides in the Decisions API documentation. You need google-genai SDK 3.2 or later. The docs' own guidance: keep each question focused on a single decision, write the full question in the instructions because the model never sees your variable names, describe every choice option and rubric level, and keep score scales linear because exponential scales skew the expected value.