Gemini Decisions API: Decision 1 price is $0.10

What Google shipped on November 11

On November 11, 2026, Min Choi and Felix Yu published the Gemini Decisions API ↗. The model id is gemini-decision-1, through the Gemini API in Google AI Studio on v1beta. Google's X post ↗ went up the same day at 18:26 UTC.

One request carries one input and up to 1,000 questions. The input mixes text, images, audio, video, and PDFs, up to 256K tokens. Decision 1 scores your options in one forward pass and returns probabilities plus a confidence score from 0 to 1. Google's post says it does not generate answer text.

Three question types: noul (yes/no, spelled that way in the post and the sample), choice (2 to 255 options), and score (an ordered rubric starting at 0). Latency is 150-250 ms end to end, median model compute 50-100 ms, for inputs up to 16K tokens, nearly flat from 1 to 100 questions. VentureBeat ↗ (Carl Franzen, November 11, 3:03 p.m. PT) says to time your own files at the full window and at 1,000 questions.

Use it for routing, tool selection, policy checks, moderation, and claims triage. Google says Notebook routes with it and reports up to 40% lower latency on everyday queries, and lists Stitch, CC's pick from 200-plus emoji, and a Search safety check for MCP tools.

Shortened from Google's breakroom sample, which also sends a photo, a choice, and a score. We did not run it.

client.decisions.create(
    model="gemini-decision-1",
    input=[{
        "type": "text",
        "text": "Breakroom counter check before the 9 AM client meeting.",
    }],
    questions={
        "has_spill": {
            "type": "noul",
            "instructions": "Is there an uncleaned liquid spill on the counter?",
        },
    },
)

The $0.10 price, and the free tier

Google prices Decision 1 at $0.10 per 1 million input tokens and $0 for output. The shared input is billed once per request. Google says that is nearly 5x cheaper, on its shared suite, than Gemini 3.5 Flash-Lite on the Interactions API.

VentureBeat's example: a 4,000-token record plus 20 checks of 50 tokens is about $0.50 per 1,000 records, against about $8.10 if each check resends the record. A later request pays for the input again.

AI Studio includes a no-cost tier. VentureBeat reports the cap at 200 requests and 50,000 input tokens per minute. The Gemini API terms ↗, fetched November 12, say unpaid services use submissions "to provide, improve, and develop Google products," and that "human reviewers may read, annotate, and process" them. Paid prompts are excluded from product improvement. Google logs paid prompts for a limited period for policy enforcement and required disclosures. In the EEA, Switzerland, and the UK, those paid data terms cover unpaid quota too. Keep customer files off the unpaid tier.

How it compares

OpenAI's Decisions API went generally available on October 22, 2026 on GPT-6 Luna. Cache reads are $0.01 per million tokens after a prefix of at least 1,024 tokens. Jev 1.14 images are opt-in PNG or JPEG, and VentureBeat states no question cap. Strands Decider 2B (October, Apache 2.0) is a text model you run yourself.

Google's launch numbers, via VentureBeat: JevBench 92.2% for Decision 1, 90.9% GPT-6 Luna, 90.0% Jev 1.14. Jev led Mind2Web Multimodal (86.8 to 86.2) and When2Call (82.0 to 79.9). Against Flash-Lite: MMAU 71.8 to 67.7, Video-MME 69.1 to 63.3. Calibration error 0.0250 versus Jev's 0.0494. Suite cost: $0.88 Decision 1, $0.68 Jev, $4.14 Flash-Lite.

VentureBeat's Artificial Analysis Decision Index v2.1 table ranks Decision 1 9th of 18 at 65.98. Milliseconds there are the index's weighted medians, separate from Google's 150-250 ms figure. The cost column estimates a million tasks from that mix.

AA-Paperclip, every answer right: Gemini 33%, Jev 35.5%, Drex Max v1.1 63%. Drex Max is $0.03 per million input tokens. Clef-Omni is $0.15 per million on Workers AI, takes audio and video, and Cloudflare publishes the weights.

Google named OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode, with Cloud support "coming soon." See Auto Code, agent frameworks, the MCP guide, Carbon /swarm, and ChatGPT vs Claude vs Gemini.

What people are saying

• "BREAKING: Something at Google is now apparently capable of making a decision!" wrote Corey Quinn ↗.

• "The scarce part isn't the 200ms latency — it's the rubric, the confidence cutoff you will actually enforce, and who owns it when a confident yes is wrong." wrote @nick_ai_agent ↗.

Reddit was not reachable for this story.

Our take

Try it now when the input is a recording, photos, or a PDF and you will enforce a cutoff. Keep the allow/deny rule in your code. AA-Paperclip's 33% against Drex Max's 63%, and the 9th-place index score, are the case for running Jev or Drex on the same labeled text first.

Use Jev for short text tickets ($0.042 per million, 92 ms on the index). VentureBeat's injection test on Jev dropped a block probability from 0.76 to 0.48 after one planted field. Use Luna when a long prefix repeats across calls. Run Strands when the text cannot leave your network. Audio and video stay on Decision 1 or Clef-Omni. When the task needs a written sentence, use a generative model such as Carbon. Send production traffic on paid quota.