Gemini Decision 1: Features, Benchmarks, Pricing, and Access

Decision models are becoming a category of their own. TypeSafe's Jev returns typed answers with probabilities instead of text, OpenAI has its own Decisions API, and on November 11, 2026, Google joined in with the Gemini Decisions API and the model behind it, Gemini Decision 1.

Gemini Decision 1 answers closed questions, never open ones: yes/no checks, multiple choice, and rubric scores, up to 1,000 of them per request. It reads text, images, audio, video, and PDFs in a single forward pass, bills zero output tokens, and costs $0.10 per 1M input tokens. Google reports 150–250 ms end-to-end latency and the lowest calibration error in its comparison, ahead of Jev.

In this article, I'll cover what Gemini Decision 1 is, what you can build with it, how it does on Google's benchmarks, and what it costs. If decision models are new to you, see our guides to TypeSafe's Jev and the OpenAI Decisions API.

To keep up with the AI and data space, I recommend subscribing to our weekly newsletter, The Median.

TL;DR

• Gemini Decision 1 is Google's first decision model: it scores options instead of writing text.
• Its edge over Jev is media: one call can judge images, audio, video, and PDFs.
• It has the best calibration in Google's chart, so its confidence scores are the ones to threshold on.
• Jev still wins on tool-call decisions and comes out cheaper across Google's shared evals.
• Use it for routing, policy checks, and grading; it can't generate a single word.

What Is Gemini Decision 1?

Gemini Decision 1 (gemini-decision-1) is a specialized multimodal model that evaluates typed questions against a shared input and returns calibrated probabilities with zero output tokens. It is the only model behind Google's new Decisions API, a separate endpoint from the Interactions API that serves Gemini's generative models.

The difference from a regular large language model (LLM) is in how the answer comes out. A generative model writes "billing" token by token, and your code parses it; Gemini Decision 1 scores every candidate option in one pass and hands back a probability for each. There is no text to parse, no thinking budget, and no output bill.

The claim Google leads with is calibration: Gemini Decision 1 has the lowest expected calibration error (ECE) in its comparison, which means its stated confidence tracks how often it is actually right. That matters more than raw accuracy when your code acts on a probability.

Gemini Decision 1 Key Features

Gemini Decision 1 does one job, so its features are about how much of that job fits into one call.

Ask up to 1,000 questions about one input in one call

You send one shared input and a map of questions, and you get one typed answer per question. A single request can carry 1 to 1,000 questions, bounded by the 262,144-token context window, and every question runs against the input in isolation, so questions can't see each other.

Three question types cover most closed decisions:

• Yes/no (noul): the probability that a statement holds, from 0.01 to 0.99.
• Choice: one option from 2 to 255, with a probability for every option.
• Score: a rating on an ordered rubric of 2 to 255 levels, returned as a probability-weighted expected level.

The practical win is billing and latency. The shared input is billed once per request, each extra question adds only its own instruction and criteria tokens, and Google says latency stays nearly flat from 1 to 100 questions. Checking a support message against 40 policy rules costs little more than checking it against one, and takes about as long.

Judge images, audio, video, and PDFs directly

Gemini Decision 1 reads media natively, so you can ask whether a scanned PDF claim matches its damage photos, or which queue a call recording belongs in, without a transcription or optical character recognition (OCR) step first. Google notes there are no separate audio transcription or OCR charges, and you can tune resolution on image and video inputs to trade visual detail against token cost.

This is the clearest gap to Jev and GPT-6 Luna, neither of which has a score on Google's audio and video benchmarks. Google's own teams use it on live media: CC from Google Labs picks an emoji reaction from more than 200 options in a fraction of a second, and Stitch's rapid canvas applies voice and tap edits almost instantly.

Two caveats from the docs. Question instructions accept text only, and the API rejects YouTube URLs and direct gs:// Cloud Storage URIs, so media goes in through the Files API, an HTTPS URL, or inline bytes.

Get a confidence score you can route on

Every answer carries a confidence score from 0 to 1, the estimated probability that the top answer is right. The intended pattern is to act automatically above a threshold and send everything else to a fallback or a human reviewer.

Google is explicit that the threshold is your job: set it from labeled examples of your own traffic and weigh false positives against false negatives. Good calibration on Google's evals is a starting point rather than proof it holds on your data.

Route requests to the right model

The headline use case is as a control plane in front of generative models. Gemini Notebook uses the Decisions API to send each incoming query to the right model tier and thinking level, serving everyday queries with up to 40% lower latency while saving deep reasoning models for complex ones.

Google Search uses it as a safety layer for Model Context Protocol (MCP) tools, detecting adversarial behavior within its latency budget. The pattern generalizes: anywhere you currently call a generative model only to get back a label, a decision model does the same job faster and without parsing.

How Does Gemini Decision 1 Perform on the Benchmarks?

Gemini Decision 1 leads 7 of the 9 accuracy benchmarks Google reports and has the lowest calibration error, but Jev 1.14 wins the two tool-and-action benchmarks and is cheaper overall. The comparison covers Jev 1.14, GPT-6 Luna on OpenAI's Decisions API, and Gemini 3.5 Flash-Lite on the Interactions API at minimal thinking. Scores are Google-reported, and the best per row is in bold.

Routing, ranking, and reasoning

The biggest lead is on New Yorker Caption (text-only), where Gemini Decision 1 scores 79.7% against Jev 1.14's 72.0% and GPT-6 Luna's 68.4%.

Most of the leads over Jev are 1–2 points, which is close to noise. When2Call is the one to watch if you're building agent routers: Jev's 2-point edge there lands on exactly the decision those routers make most often.

Audio and video

Gemini Decision 1 scores 69.1% on Video-MME without subtitles, against 63.3% for Gemini 3.5 Flash-Lite. Jev and GPT-6 Luna have no score on either media benchmark.

So on audio and video, the only comparison is against Google's own generative model. Beating Flash-Lite by about 4–6 points while billing no output tokens is still the stronger case for media decisions.

Calibration and cost

Gemini Decision 1's Top-1 ECE of 0.0250 across the 7 shared evals is the best in the chart, ahead of Jev 1.14 at 0.0494 and GPT-6 Luna at 0.0721.

Price is where the picture flips. Across the same 7 evals, Jev is the cheapest model in the chart.

Google's "nearly 5x cheaper" claim is against its own Gemini 3.5 Flash-Lite. Against Jev, you're paying about 30% more for better calibration and media support.

Gemini Decision 1 Pricing and Availability

Gemini Decision 1 costs $0.10 per 1M input tokens on the paid tier, with no output charge because it generates zero output tokens. Text, image, video, audio, and PDF input all bill at the same rate.

• Free tier: available in Google AI Studio for prototyping, at 200 requests per minute and 10,000 requests per day. Free-tier data is used to improve Google's products; paid-tier data is not.
• Paid tiers: Tier 1 allows 1,500 requests per minute, rising to 15,000 at Tier 3.
• Not supported: context caching, the Batch API, and Flex or Priority inference, so repeated input across requests is billed in full every time.
• Status: available now on the Gemini Developer API (v1beta).

The missing context cache matters more than it looks. If you check the same long document with a new question every few seconds, you pay for that document on every call; batch the questions into one request instead.

How to Get Access to Gemini Decision 1?

The API model ID is gemini-decision-1, called through client.decisions.create in the google-genai SDK with the v1beta API version. Google AI Studio also has a Decisions Playground for testing questions before you write code.

On OpenRouter, the model is google/gemini-decision-1, served through the System One endpoint used for Jev, so not the chat completions endpoint. Google says integrations are rolling out across LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode, with support for Gemini in Google Cloud coming soon. It is not listed in the Vertex AI model docs as of November 12, 2026.

Here's a minimal expense-claim check that asks two questions about one note:

from google import genai

client = genai.Client(http_options={"api_version": "v1beta"})
result = client.decisions.create(
    model="gemini-decision-1",
    input="Team dinner, 6 people, $412 incl. wine. Receipt attached.",
    questions={
        "has_alcohol": {"type": "noul", "instructions": "Does the claim include alcohol?"},
        "category": {
            "type": "choice",
            "instructions": "Which expense category fits this claim?",
            "criteria": {"meals": "Food and drink.", "travel": "Flights, trains, hotels.", "other": "Anything else."},
        },
    },
)
print(result.answers["has_alcohol"].noul, result.answers["category"].choice)Powered by

For a full worked example of a decision-model pipeline, our Jev ticket router tutorial covers the same pattern of questions, thresholds, and fallbacks.

Final Thoughts

Gemini Decision 1 makes decision models a three-vendor category, and Google brought the one thing TypeSafe and OpenAI didn't: a decision model that reads images, audio, video, and PDFs natively. With the best calibration in its own chart, it's the safer pick when you act on confidence scores.

It isn't a clean win. Jev 1.14 is cheaper across Google's shared evals and better on tool-call decisions, so a text-only agent router has no urgent reason to switch. Anything involving media or PDFs is where I'd start with Gemini Decision 1.

If you're keen to build applications around models like this, I recommend checking out our Developing AI Applications skill track.