Gemini Decision 1
Google Gemini Decision 1
Google's proprietary multimodal decision model with zero output tokens, a 256K context window, $0.10/M input pricing ($0 output), and a 65.98 score on the AA Decision Index v2.1.
At a glance
- Context window
- 256K tokens
- Max output
- —
- Text input / output · per 1M tokens
- $0.100 / Free
- Knowledge cutoff
- —
Benchmark snapshot
Best result in each category
General EvaluationAA Decision Index v2.1
66Normal (No Tools) · #4 of 5
Benchmark details
1 results · Best published score per benchmark, grouped by category
| Benchmark · setting | Score | Rank |
|---|---|---|
| General Evaluation | ||
| AA Decision Index v2.1 Normal (No Tools) · Capability Indices | 66 | 4 / 5 |
Rank is this model's position among every model with a published result on that leaderboard; the bar shows how far up the field that is. Orange marks first place on leaderboards with at least 5 models. Unranked leaderboards publish a composite or direction-less score.
Back to benchmark snapshot ↑API pricing
Text rates in USD per million tokens; other billing units shown below · effective 2026-11-11
- Input
- $0.10
- Output
- $0.00
Blended price4:1 input:output · same basis as the price ranking
$0.08/ 1M tokens
Image pricing
- image input$0.10 / 1M tokens
Official Gemini Developer API paid-tier image-input price ($0.10/1M tokens).
Specifications
- Provider
- Google Deep Mind
- Model type
- Multimodal model
- Released
- November 11, 2026
- Architecture
- Undisclosed
- Context window
- 256K tokens
- Input
- Text, Image, Audio, Video
- Output
- Text
- Weights
- Closed / API only
- Commercial use
- Proprietary
Model analysis
Gemini Decision 1 (API ID gemini-decision-1) is a proprietary multimodal decision model released by Google DeepMind on November 11, 2026 to power the new Gemini Decisions API. Rather than decoding answer text token by token, it evaluates a shared input in a single forward pass and returns calibrated probabilities and confidence scores. Google highlights top results across nine launch benchmarks, the lowest calibration error, and nearly 5x lower cost than generative Gemini 3.5 Flash-Lite; however, on the independent 26-model Artificial Analysis Decision Index v2.1, Gemini Decision 1 scores 65.98 (16th overall) at an effective cost of $345.41 per 1M tasks—substantially more expensive than several higher-scoring dedicated decision models.
Single-pass scoring replaces token decoding, within closed-set bounds
Unlike chat or reasoning LLMs, Gemini Decision 1 only evaluates developer-defined closed-set questions. A single request shares a 262,144-token (256K) context window across input and up to 1,000 independent questions, returning structured outputs across three types: noul (yes/no probability from 0.01 to 0.99 plus confidence), choice (2 to 255 fixed options with normalized per-option probabilities), and score (2 to 255 ordered rubric levels returning a weighted expected score and per-level probabilities).
The shared input natively accepts text, images, audio, video, and PDF documents (inline base64 up to 100 MB per request or 50 MB for PDFs; Files API URIs or public/signed HTTPS URLs up to 15 MB). Because it generates zero text tokens, platform modality metadata maps its output format to text for schema compatibility, while its actual output token limit is 0. Official documentation notes key restrictions: no generateContent chat endpoint, Thinking mode, function calling, general Structured Outputs, context caching, Batch API, Flex/Priority inference, YouTube URLs, direct gs:// URIs, media inside question instructions, or multi-turn conversation/thought/tool-call blocks. Parameter count, training knowledge cutoff, and a full supported-language list remain undisclosed.
Headline $0.10/M input pricing masks higher per-task costs in third-party tests
Google prices Gemini Decision 1 at $0.10 per 1M input tokens across text, image, audio, video, and PDFs, with $0 output tokens. Shared context is billed once per request, and each question adds only its own instruction and criteria tokens. A no-cost prototyping tier is available in Google AI Studio (200 RPM, 50,000 TPM, 10,000 RPD; free-tier data may be used to improve Google products), while paid Tiers 1–3 scale up to 15,000 RPM and 10,000,000 TPM.
The nominal per-token rate looks inexpensive, yet relative cost depends heavily on the baseline chosen. Across seven shared evaluations in Google's launch post, total spend was $0.88 versus $4.14 for generative Gemini 3.5 Flash-Lite (minimal thinking via the Interactions API)—a 4.70x savings. In that same Google table, however, dedicated decision model TypeSafe Jev 1.14 cost $0.68 total, 22.7% less than Gemini Decision 1. Under Artificial Analysis Decision Index v2.1's real task billing (read November 2026), Gemini Decision 1 averages $345.41 per 1M tasks, which is 2.21x costlier than TypeSafe Jev 1.14 ($156.23/1M), 3.36x costlier than top-ranked decision model Drex Max v1.1 ($102.68/1M), and 5.08x costlier than Decider v1.3 ($68.04/1M). On latency, Google claims 150–250 ms end-to-end (50–100 ms compute) for contexts under 16K tokens, and Artificial Analysis records a 211 ms median per task; on OpenRouter (single provider: Google AI Studio, 99.97% 3-day availability), observed P50 end-to-end latency over the past week is 0.39 s (390 ms).
On a 26-model independent index, overall rank drops to mid-pack
Google's November 11, 2026 announcement compared Gemini Decision 1 against three baselines (TypeSafe Jev 1.14, GPT-6 Luna OpenAI Decisions API, and Gemini 3.5 Flash-Lite) across nine benchmarks (full table below). It led seven of nine, including audio benchmark MMAU (71.8%) and subtitle-free Video-MME (69.1%) where the other decision models had no reported scores, and posted the lowest Top-1 Expected Calibration Error across the seven shared text/image evals at 0.0250 (vs. 0.0494 for Jev 1.14 and 0.0721 for GPT-6 Luna Decisions). Still, Jev 1.14 edged it out on Mind2Web Multimodal (86.8% vs. 86.2%) and When2Call (82.0% vs. 79.9%).
In Artificial Analysis Decision Index v2.1 across 26 models (22 dedicated decision models plus 4 generative LLM baselines), Gemini Decision 1 scores 65.98 overall (Accuracy Index 45.53, Brier score 0.1752), placing 16th overall and 12th among dedicated decision models—behind leaders Drex Max v1.1 (72.62, Brier 0.1600), SubQ Snap 1.2 (72.58, 162 ms), GLiDE v1.2 (72.52), Mammoth 27B (71.36, 106 ms), d1-ultra (70.58), and Decider v1.3 (70.41), and just behind TypeSafe Jev 1.14 (66.23, 92 ms). Its composite score is dragged down by document extraction/decisions on 30%-weighted AA-Paperclip (56.83, 33% all-pass rate vs. 72.82 for Drex Max v1.1) and tool-calling on When2Call (80.50, 75.20% accuracy vs. 96.10 for SubQ Snap 1.2). Conversely, on hard reasoning and factual abstention it leads all 22 dedicated decision models, ranking first among them on HLE-MultiChoice v1.1 (48.71 Decision Score, Brier 0.1968) and AA-Omniscience-Decision v1.1 (66.28, 44.80% accuracy, 49.64% non-hallucination rate).
The complete launch comparison table below is reproduced from Google's November 11, 2026 announcement; refusals, safety blocks, and unaccepted requests count as incorrect with the full dataset as denominator, and scores are Top-1 Accuracy unless noted:
| Category | Benchmark | Protocol / Notes | Gemini Decision 1 (Decisions API) | TypeSafe Jev 1.14 | GPT-6 Luna (OpenAI Decisions API) | Gemini 3.5 Flash-Lite (Interactions API minimal) |
|---|---|---|---|---|---|---|
| Key benchmarks | JevBench | Public split | 92.2% | 90.0% | 90.9% | 86.6% |
| Key benchmarks | Amazon ESCI | Macro-F1 | 58.7% | 56.7% | 54.7% | 47.2% |
| Key benchmarks | New Yorker Caption | Text-only | 79.7% | 72.0% | 68.4% | 73.7% |
| Key benchmarks | Mind2Web Multimodal | — | 86.2% | 86.8% | 81.8% | 78.0% |
| Key benchmarks | PubMedQA | — | 79.8% | 78.2% | 74.4% | 77.6% |
| Key benchmarks | Global-MMLU-Lite | Non-English | 86.0% | 85.0% | 79.6% | 78.8% |
| Key benchmarks | When2Call | — | 79.9% | 82.0% | 70.4% | 60.9% |
| Key benchmarks | MMAU | — | 71.8% | — | — | 67.7% |
| Key benchmarks | Video-MME | No subtitles | 69.1% | — | — | 63.3% |
| Calibration | Top-1 ECE | 7 shared evals, lower is better | 0.0250 | 0.0494 | 0.0721 | 0.2104 |
| Cost | Total price | 7 shared evals | $0.88 | $0.68 | $0.93 | $4.14 |
Strongest on native audio/video and hard knowledge checks; text routing has cheaper alternatives
Gemini Decision 1 is the clearest fit when your shared input includes audio, video, or visually dense PDFs and you need to evaluate dozens or hundreds of compliance checks, guardrails, or rubric dimensions in a single low-latency call—or when calibrated abstention on knowledge-intensive choices matters, within the Gemini Developer API / AI Studio / OpenRouter ecosystem (with LangChain, LiteLLM, Vercel AI SDK, and Pydantic AI integrations). For pure-text routing, document triage, or tool-call gating, higher-ranked and cheaper dedicated alternatives exist: Drex Max v1.1 (72.62, $102.68/1M tasks), SubQ Snap 1.2 (72.58, $136.90/1M tasks, 162 ms), and Decider v1.3 (70.41, $68.04/1M tasks) deliver higher Decision Index scores at roughly 2.5x to 5.1x lower per-task cost, while TypeSafe Jev 1.14 (92 ms, $156.23/1M) and Decider Flash v1.3 (178 ms, $34.02/1M) target ultra-low latency or minimal spend. Where open-ended extraction or multi-step reasoning justifies 3–32 second latency, generative baselines like Claude Haiku 5.5 (Max: 76.49) and GPT-6 Luna (Max: 75.67) still lead on raw accuracy.
Sources: Google Announcement (2026-11-11); Google AI Studio Model Spec; Decisions API Docs; Pricing & Rate Limits; Artificial Analysis Decision Index v2.1; OpenRouter Model Page.
Last reviewed 2026-11-15
Questions
- How is Gemini Decision 1 different from standard Gemini models?
- Gemini Decision 1 does not generate text token-by-token and does not support the generateContent chat endpoint or Thinking mode. Through the Decisions API, it evaluates up to 1,000 structured yes/no (noul), single-select (choice, 2–255 options), and ordered rubric (score, 2–255 levels) questions in one forward pass and returns calibrated probabilities and confidence scores with 0 output tokens.
- What does Gemini Decision 1 cost?
- Paid-tier pricing is $0.10 per 1M input tokens across text, image, audio, video, and PDF inputs, with $0 output token charges. Shared input is billed once per request regardless of question count. Google AI Studio also provides a free prototyping tier (200 RPM, 50,000 TPM, 10,000 RPD; free-tier inputs may be used for product improvement). Batch API and context caching are not supported.
- How does Gemini Decision 1 perform on independent benchmarks?
- On the 26-model Artificial Analysis Decision Index v2.1, Gemini Decision 1 scores 65.98 overall (16th overall, 12th among dedicated decision models) with a 211 ms median latency and $345.41 cost per 1M tasks. It ranks first among dedicated decision models on HLE-MultiChoice v1.1 (48.71) and AA-Omniscience-Decision v1.1 (66.28), while trailing leaders such as Drex Max v1.1 (72.62) and SubQ Snap 1.2 (72.58) on AA-Paperclip and When2Call.
Gemini Decision 1 vs competitors
Comparable results only · record is Gemini Decision 1's wins–losses
| Benchmark | Gemini Decision 1This model | vs Jev 1.14Record 0–1 | vs Microsoft-Decision-1Record 1–0 | vs Claude Haiku 5.5Record 0–1 | vs GPT-6 LunaRecord 0–1 |
|---|---|---|---|---|---|
| AA Decision Index v2.1 | 66 | 66.2 | 62.1 | 76.5 | 75.7 |
Primary sources
- Gemini Decision 1 technical reportblog.google
- Gemini Decision 1 online demoaistudio.google.com
Community discussion
Share the API, setup, and tradeoffs you actually tested. Rate the model once; comment as often as you like.
Used this model? Sign in with Google to rate it and share what worked, what did not, and which API or setup you tested.
No community comments yet.
