Skip to content
System One

Gemini Decision 1

Gemini Decision 1 is Google's hosted decision model, served by the new Decisions API in the Gemini API. You send text, images, audio, video or PDFs with up to 1,000 typed questions, and it returns a probability for every answer. Released on 11 November 2026 at $0.10 per million input tokens, with output free.

Updated

What Gemini Decision 1 is

Gemini Decision 1 is Google’s decision model, a System One model that answers typed questions instead of writing text. Google announced it on 11 November 2026 and opened it the same day. It runs behind the Gemini Decisions API, a new endpoint in the Gemini API in Google AI Studio, under the model id gemini-decision-1. It scores the options in one forward pass, with no token-by-token decoding.

What it returns

You call POST https://generativelanguage.googleapis.com/v1beta/decisions with one shared input and a map of named questions. The input can mix text, images, audio, video and PDFs. A Noul returns the probability that a statement holds. A Choice picks one of 2 to 255 options and gives a probability for each. A Score rates the input on up to 255 ordered levels and returns a weighted score with per-level probabilities. Every answer has a confidence from 0 to 1. One request takes up to 1,000 questions over 262,144 input tokens.

Price and speed

Google charges $0.10 per million input tokens, with free output and a free tier in AI Studio. The shared input is billed once per request, so 50 questions about one PDF do not pay for the PDF 50 times. Google’s own latency figures are 150 to 250 ms end to end for inputs up to 16K tokens, nearly flat from 1 to 100 questions.

What Google reports

Every quality number in Google’s launch table is vendor-run. Google compared the model with three others: Jev 1.14, GPT-6 Luna on the OpenAI Decisions API and Gemini 3.5 Flash-Lite. Gemini Decision 1 scored highest of the four on five of the seven shared evals, such as 92.2 on the JevBench public split against 90.0 for Jev. Jev scored higher on Mind2Web Multimodal and When2Call, and the seven evals cost $0.68 on Jev against $0.88 on Gemini Decision 1.

Independent results

On the Artificial Analysis Decision Index version 2.1, checked on 12 November 2026, it scores 66.0. Jev 1.14 is at 66.2 and GPT-6 Luna (Decisions) at 60.2; Drex Max v1.1 and SubQ Snap 1.2 are highest at 72.6. Its calibration error, which this site computed from AA’s published bins, is 0.013 against 0.054 for Jev. AA measured $345 per million tasks and 211 ms per task, against $156 and 92 ms for Jev. JevBench added it in release v1.8.10, measured on 11 November 2026, with a Capability Score of 78.5, 13th on the hosted API board, against 75.0 for Jev 1.14.

What it’s for

Google names routing, policy checks across many rules at once, rubric scoring, tool selection and classification over live media. It says it already uses the model for model routing in Gemini Notebook and as a safety layer for MCP tools in Google Search. It does not generate text or call functions, and the Batch API is not supported.

Access

It is live in the Gemini API with a Gemini API key, through google-genai for Python and @google/genai for JavaScript, version 3.2 or later. OpenRouter lists it as google/gemini-decision-1 at the same price. Google says Google Cloud support is coming soon.

Specifications

System One Score91.5 from 2 of 3 boardsArtificial Analysis Decision Index 90.9, JevBench 92.1. How the score works
InputsText, images, audio, video, PDFsUp to 100 MB inline per request, or 50 MB for PDFs.
Question typesChoiceScoreNoul
Max Choice options255
Score levelsUp to 255
Questions per call1,000
Total context262,144 tokens
State budgetNot documented
Rate limitGoogle's rate limits page, checked 2026-11-12: free tier 200 requests and 50,000 tokens per minute, 10,000 requests per day; Tier 1 1,500 requests and 1,000,000 tokens per minute, 50,000 per day; Tier 2 5,000 requests and 4,000,000 tokens per minute; Tier 3 15,000 requests and 10,000,000 tokens per minute. Tiers 2 and 3 have no daily request cap.
EndpointPOST https://generativelanguage.googleapis.com/v1beta/decisions
LicenceNot published as of 12 Nov 2026
SDKsPython: google-genaiJavaScript: @google/genai

Google's model page gives an input limit of 262,144 tokens shared by up to 1,000 questions per request; the launch article rounds it to 256K. Choice takes 2 to 255 options and Score 2 to 255 ordered levels. Question keys match [A-Za-z0-9_-]+, up to 64 characters. Inline media is capped at 100 MB per request (50 MB for PDFs) and external URLs at 15 MB each. Image and video resolution can be set per block. Google publishes no separate state-token limit. The Batch API is not supported. Checked on 2026-11-12.

Versions

  • gemini-decision-1, 11 Nov 2026, Announced and available on 11 November 2026 in the Gemini API, on the v1beta endpoint. Google's model page lists gemini-decision-1 as the only version. OpenRouter lists it as google/gemini-decision-1. Google says Google Cloud support is coming soon. Release notes

Use cases

What people use Gemini Decision 1 for, one page per pattern.

Workflow controlIntermediate

Intent and model routing with System One models

One Jev call reads an incoming request and returns its intent as a label plus a difficulty rating on a scale you wrote. Your router reads both numbers and picks the handler: deterministic code, a cheap model, an expensive one, or a human queue.

ChoiceScore
Safety and qualityIntermediate

LLM guardrails with System One models

Put one Jev request in front of an LLM and one behind it. Yes/no questions return the probability that each hazard holds, a Score rates how much harm complying would do, and your thresholds turn those numbers into pass, review, block, or a crisis path.

NoulScore
Real-time and agentsIntermediate

Agent routing and skill selection with System One models

An agent choosing from a long skill roster reads one truncated line per entry and often loads the wrong thing. Jev ranks every entry in one request and separately answers whether any skill applies at all, so the agent gets a short hint instead of a guess.

ChoiceNoul
Retrieval and knowledgeIntermediate

Semantic reranking with System One models

Keyword or vector search narrows thousands of documents to a shortlist but rarely puts the right one first. Ask Jev one yes/no question about each query and candidate pair, take the probability it returns as the score, and sort the shortlist by it, highest first.

Noul
Real-time and agentsIntermediate

Real-time control with System One models

A model inside a control loop has to answer before the next frame. Jev returns a typed decision rather than text, and TypeSafe reports end to end latency of 70ms to 500ms, which puts a semantic judgment inside the budget of a game tick or an interface response.

ChoiceNoul

Benchmarks

Published scores, each linked to its source. A row marked vendor-run was measured by the vendor itself, so read it as one side's result. The compare hub sets these beside the other models.

BenchmarkMetricScoreRun byDate
Artificial Analysis Decision Index 2.1Index score66 scoreIndependent12 Nov 2026
Artificial Analysis Decision Index 2.1Expected calibration error0.0132 ECEIndependent12 Nov 2026
Artificial Analysis Decision Index 2.1Cost per 1M tasks345 USDIndependent12 Nov 2026
Artificial Analysis Decision Index 2.1Time per task211 msIndependent12 Nov 2026
Amazon ESCImacro-F158.7 scorevendor-run11 Nov 2026
Global-MMLU-Lite (non-English)Top-1 accuracy86 %vendor-run11 Nov 2026
Google launch suite (7 shared evals)Top-1 expected calibration error0.025 ECEvendor-run11 Nov 2026
Google launch suite (7 shared evals)Total price0.88 USDvendor-run11 Nov 2026
JevBench 1.8.10Capability Score78.5 scoreIndependent11 Nov 2026
JevBench (public split)Top-1 accuracy92.2 %vendor-run11 Nov 2026
Mind2Web MultimodalTop-1 accuracy86.2 %vendor-run11 Nov 2026
MMAUTop-1 accuracy71.8 %vendor-run11 Nov 2026
New Yorker Caption (text only)Top-1 accuracy79.7 %vendor-run11 Nov 2026
PubMedQATop-1 accuracy79.8 %vendor-run11 Nov 2026
Video-MME (no subtitles)Top-1 accuracy69.1 %vendor-run11 Nov 2026
When2CallTop-1 accuracy79.9 %vendor-run11 Nov 2026
  • Artificial Analysis Decision Index, Index score: Row: Gemini Decision 1, host Google AI Studio. Raw value 65.977. Jev 1.14 scores 66.2 and GPT-6 Luna (Decisions) 60.2 on the same version. Sub-scores: When2Call accuracy 75.2, Omniscience accuracy 44.8, HLE-MultiChoice accuracy 13.2. Version 2.1 figures do not compare with version 1.0 figures.
  • Artificial Analysis Decision Index, Expected calibration error: Row: Gemini Decision 1. Computed by this site from AA's published calibration bins (all answers pooled); AA does not publish an ECE. Lower is better. Same method on version 2.1: Jev 1.14 0.0536, GPT-6 Luna (Decisions) 0.1103.
  • Artificial Analysis Decision Index, Cost per 1M tasks: Row: Gemini Decision 1. AA's measured cost per million tasks on its version 2.1 evals. Jev 1.14 $156, GPT-6 Luna (Decisions) $339.
  • Artificial Analysis Decision Index, Time per task: Row: Gemini Decision 1. Time per task. Jev 1.14 92 ms, GPT-6 Luna (Decisions) 206 ms.
  • Amazon ESCI, macro-F1: Google's launch table. Jev 1.14 56.7, GPT-6 Luna 54.7, Gemini 3.5 Flash-Lite 47.2.
  • Global-MMLU-Lite (non-English), Top-1 accuracy: Google's launch table. Jev 1.14 85.0, GPT-6 Luna 79.6, Gemini 3.5 Flash-Lite 78.8.
  • Google launch suite (7 shared evals), Top-1 expected calibration error: Google's launch table, computed by Google. Lower is better. Jev 1.14 0.0494, GPT-6 Luna 0.0721, Gemini 3.5 Flash-Lite 0.2104.
  • Google launch suite (7 shared evals), Total price: Google's launch table. Jev 1.14 $0.68, GPT-6 Luna $0.93, Gemini 3.5 Flash-Lite $4.14.
  • JevBench, Capability Score: Row: Gemini Decision 1, measured 2026-11-11. Intelligence 67.4, Calibration 89.6. 13th by Capability Score on the hosted API board. First listed in release v1.8.10. Board release v1.8.10, read 2026-11-15.
  • JevBench (public split), Top-1 accuracy: Google's launch table, refusals scored incorrect. Jev 1.14 90.0, GPT-6 Luna (OpenAI Decisions API) 90.9, Gemini 3.5 Flash-Lite (Interactions API, minimal) 86.6. Not a run on the JevBench board.
  • Mind2Web Multimodal, Top-1 accuracy: Google's launch table. Jev 1.14 86.8, GPT-6 Luna 81.8, Gemini 3.5 Flash-Lite 78.0.
  • MMAU, Top-1 accuracy: Google's launch table, audio. Gemini 3.5 Flash-Lite 67.7; no score for Jev 1.14 or GPT-6 Luna in Google's table.
  • New Yorker Caption (text only), Top-1 accuracy: Google's launch table. Jev 1.14 72.0, GPT-6 Luna 68.4, Gemini 3.5 Flash-Lite 73.7.
  • PubMedQA, Top-1 accuracy: Google's launch table. Jev 1.14 78.2, GPT-6 Luna 74.4, Gemini 3.5 Flash-Lite 77.6.
  • Video-MME (no subtitles), Top-1 accuracy: Google's launch table, video. Gemini 3.5 Flash-Lite 63.3; no score for Jev 1.14 or GPT-6 Luna in Google's table.
  • When2Call, Top-1 accuracy: Google's launch table. Jev 1.14 82.0, GPT-6 Luna 70.4, Gemini 3.5 Flash-Lite 60.9.