The Gemini 3 Pro model card is an eight-page document published by Google DeepMind with the launch of Gemini 3 Pro in November 2025; the version summarised here is marked "Model Release: November 2025, Last Updated: May 2026" (Source: https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-Pro-Model-Card.pdf). It is the base card for the Gemini 3 family: later cards, including the Gemini 3.1 Pro model card, refer back to it for architecture, training data and safety policies. The card states that it "includes more essential information about the Gemini 3 family of models than previous model cards did", covering the training dataset, distribution and intended uses. A companion page gives the evaluation methodology (Source: https://deepmind.google/models/evals-methodology/gemini-3-pro/), and a separate Frontier Safety Framework report gives the dangerous-capability evaluations in detail.
Model information
Gemini 3 Pro is described as a natively multimodal reasoning model that accepts text, images, audio and video with a context window of up to 1M tokens and produces text with a 64K-token output, with an optional Deep Think mode for complex problem-solving at inference time. The card states that it "is not a modification or a fine-tune of a prior model", and that later members of the family are based on it; the May 2026 version lists Gemini 3 Pro Image, Gemini 3 Flash, Gemini 3.1 Pro, Gemini 3.1 Flash Image, Gemini 3.1 Flash-Lite, Gemini 3.1 Flash Live and Gemini 3.5 Flash. The architecture is a sparse mixture-of-experts transformer with native multimodal support for text, vision and audio inputs. Parameter counts are not given. The knowledge cutoff is January 2025.
The pre-training data is described as publicly available web documents, text, code, images, audio and video; post-training used instruction-tuning, reinforcement-learning and human-preference data, with reinforcement learning on multi-step reasoning, problem-solving and theorem-proving data. The card lists the data sources as publicly downloadable datasets, crawled data, commercially licensed data, user data from Google products "in accordance with Google's relevant terms of service, privacy policy, service-specific policies, and pursuant to user controls", data Google acquires or generates in its business or from its workforce, and AI-generated synthetic data. Filtering included deduplication, honoring robots.txt, safety filtering and quality filtering. Training ran on Google's Tensor Processing Units using JAX and ML Pathways. Distribution channels are the Gemini app, Google Cloud / Vertex AI, Google AI Studio, the Gemini API, Google AI Mode and Google Antigravity.
Capability evaluation
The results table, labelled "as of November, 2025", compares Gemini 3 Pro with Gemini 2.5 Pro, Claude Sonnet 4.5 and GPT-5.1 on Humanity's Last Exam, ARC-AGI-2, GPQA Diamond, AIME 2025, MathArena Apex, MMMU-Pro, ScreenSpot-Pro, CharXiv Reasoning, OmniDocBench 1.5, Video-MMMU, LiveCodeBench Pro, Terminal-Bench 2.0, SWE-Bench Verified, τ2-bench, Vending-Bench 2, the FACTS Benchmark Suite, SimpleQA Verified, MMMLU, Global PIQA and MRCR v2 (8-needle). The card states that Gemini 3 Pro "significantly outperforms Gemini 2.5 Pro across a range of benchmarks requiring enhanced reasoning and multimodal capabilities". Gemini 3 Pro has the highest figure in every row except SWE-Bench Verified, where Claude Sonnet 4.5 is highest; on AIME 2025 with code execution Gemini 3 Pro and Claude Sonnet 4.5 are both listed at 100%. In the PDF the table is an embedded image.
The methodology page states that all Gemini scores are pass@1, that "single attempt" settings allow no majority voting or parallel test-time compute, and that runs used the Gemini API model gemini-3-pro-preview with default sampling. Figures for non-Gemini models are the providers' self-reported numbers at high reasoning settings, except that Google DeepMind computed MMMU-Pro, ScreenSpot-Pro, CharXiv Reasoning, OmniDocBench 1.5, Video-MMMU, MMMLU and Global PIQA itself through the providers' APIs. Third-party figures are used for ARC-AGI-2 (ARC Prize Verified, semi-private set), MathArena Apex (matharena.ai), LiveCodeBench Pro and Terminal-Bench 2.0 (public leaderboards), Vending-Bench 2 (Andon Labs) and SimpleQA Verified (the Kaggle leaderboard). The ScreenSpot-Pro figure for Gemini 3 Pro uses function calling with a screenshot-capture tool at "extra_high" media resolution; the page gives a lower figure at "high" resolution. The τ2-bench figure is the average of the Retail, Airline and Telecom domains, whose individual scores the page also gives. Google notes that the FACTS Benchmark Suite results are not comparable with its earlier FACTS Grounding results.
Content safety
The card lists Google's safety policies for Gemini (child sexual abuse material, hate speech, dangerous content, harassment, sexually explicit content, and medical advice contrary to consensus) and four evaluation types: training and development evaluations, human red teaming by specialist teams outside the development team, automated red teaming, and ethics and safety reviews. It reports five internal automated evaluations as absolute percentage changes relative to Gemini 2.5 Pro: text-to-text safety, multilingual safety, image-to-text safety, tone, and unjustified refusals. Google states that Gemini 3 Pro "outperforms Gemini 2.5 Pro across both safety and tone, while keeping unjustified refusals low", that the results are not comparable with earlier cards because the evaluations changed, and that manual review found losses were "overwhelmingly either a) false positives or b) not egregious". Evaluations using Deep Think mode are described as consistent with the original assessment. No absolute violation or refusal rates are given.
On human red teaming, the card says Gemini 3 Pro satisfied required launch thresholds for child safety, showed similar or improved safety performance compared with Gemini 2.5 Pro, and that an expanded red-teaming scope "found no egregious concerns". It names two main risks: jailbreak vulnerability, described as improved over Gemini 2.5 Pro "but still an open research problem", and possible degradation in multi-turn conversations. Mitigations listed are dataset filtering, conditional pre-training, supervised fine-tuning, reinforcement learning from human and critic feedback, safety policies and desiderata, and product-level safety filtering.
Frontier safety
The card summarises an evaluation under the September 2025 Frontier Safety Framework and reports that Gemini 3 Pro reached no critical capability level (CCL). For CBRN, the model "provides accurate and occasionally actionable information" but generally not novel or sufficiently complete instructions. For cybersecurity, it solved 11 of 12 hard challenges on the v1 key-skills benchmark and none of 13 v2 challenges end-to-end; the alert threshold was met. For harmful manipulation, efficacy improved on a non-generative-AI baseline with no significant uplift over prior models. For machine learning R&D, the model did better than Gemini 2.5 models on RE-Bench, especially on the Scaling Law Experiment and Optimize LLM Foundry tasks, with an aggregate score "substantially below the alert threshold". For misalignment (exploratory), the agent solved 3 of 11 situational-awareness challenges and 1 of 4 stealth challenges. Deep Think mode results are described as consistent with these.
Reported results
Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.
| Evaluation | Model | Score | Setting | Section |
|---|---|---|---|---|
| OmniDocBench 1.5 — Overall Edit Distance (lower is better) | Gemini 3 Pro | 0.115 | average Edit Distance across Text, Formula, Table and ReadingOrder sub-metrics, official OmniDocBench code and data, following the DeepSeek-OCR methodology (arXiv 2510.18234); pass@1, Gemini API model-id gemini-3-pro-preview, default sampling; developer-reported | Evaluation – Results (table, PDF p. 5) |
| OmniDocBench 1.5 — Overall Edit Distance (lower is better) | Gemini 2.5 Pro | 0.145 | average Edit Distance across Text, Formula, Table and ReadingOrder sub-metrics, official OmniDocBench code and data, following the DeepSeek-OCR methodology (arXiv 2510.18234); developer-reported (Google DeepMind, in the Gemini 3 Pro model card) | Evaluation – Results (table, PDF p. 5) |
| OmniDocBench 1.5 — Overall Edit Distance (lower is better) | Claude Sonnet 4.5 | 0.145 | average Edit Distance across Text, Formula, Table and ReadingOrder sub-metrics, official OmniDocBench code and data, following the DeepSeek-OCR methodology (arXiv 2510.18234); reported by Google DeepMind in the Gemini 3 Pro model card; computed by Google DeepMind through the provider's official API | Evaluation – Results (table, PDF p. 5) |
| OmniDocBench 1.5 — Overall Edit Distance (lower is better) | GPT-5.1 | 0.147 | average Edit Distance across Text, Formula, Table and ReadingOrder sub-metrics, official OmniDocBench code and data, following the DeepSeek-OCR methodology (arXiv 2510.18234); reported by Google DeepMind in the Gemini 3 Pro model card; computed by Google DeepMind through the provider's official API | Evaluation – Results (table, PDF p. 5) |
| Video-MMMU | Gemini 3 Pro | 87.6% | media_resolution=HIGH (280 tokens per frame), temperature 0; pass@1, Gemini API model-id gemini-3-pro-preview, default sampling; developer-reported | Evaluation – Results (table, PDF p. 5) |
| Video-MMMU | Gemini 2.5 Pro | 83.6% | media_resolution=HIGH (280 tokens per frame), temperature 0; developer-reported (Google DeepMind, in the Gemini 3 Pro model card) | Evaluation – Results (table, PDF p. 5) |
| Video-MMMU | Claude Sonnet 4.5 | 77.8% | reported by Google DeepMind in the Gemini 3 Pro model card; computed by Google DeepMind through the provider's official API | Evaluation – Results (table, PDF p. 5) |
| Video-MMMU | GPT-5.1 | 80.4% | reported by Google DeepMind in the Gemini 3 Pro model card; computed by Google DeepMind through the provider's official API | Evaluation – Results (table, PDF p. 5) |
| FACTS Benchmark Suite | Gemini 3 Pro | 70.5% | held-out internal grounding, parametric, multimodal and search-retrieval benchmarks; not comparable with earlier FACTS Grounding results; pass@1, Gemini API model-id gemini-3-pro-preview, default sampling; developer-reported | Evaluation – Results (table, PDF p. 5) |
| FACTS Benchmark Suite | Gemini 2.5 Pro | 63.4% | held-out internal grounding, parametric, multimodal and search-retrieval benchmarks; not comparable with earlier FACTS Grounding results; developer-reported (Google DeepMind, in the Gemini 3 Pro model card) | Evaluation – Results (table, PDF p. 5) |
| FACTS Benchmark Suite | Claude Sonnet 4.5 | 50.4% | held-out internal grounding, parametric, multimodal and search-retrieval benchmarks; not comparable with earlier FACTS Grounding results; reported by Google DeepMind in the Gemini 3 Pro model card; provider's self-reported number (high reasoning) unless stated otherwise | Evaluation – Results (table, PDF p. 5) |
| FACTS Benchmark Suite | GPT-5.1 | 50.8% | held-out internal grounding, parametric, multimodal and search-retrieval benchmarks; not comparable with earlier FACTS Grounding results; reported by Google DeepMind in the Gemini 3 Pro model card; provider's self-reported number (high reasoning) unless stated otherwise | Evaluation – Results (table, PDF p. 5) |
| Global PIQA | Gemini 3 Pro | 93.4% | commonsense reasoning across 100 languages and cultures; pass@1, Gemini API model-id gemini-3-pro-preview, default sampling; developer-reported | Evaluation – Results (table, PDF p. 5) |
| Global PIQA | Gemini 2.5 Pro | 91.5% | commonsense reasoning across 100 languages and cultures; developer-reported (Google DeepMind, in the Gemini 3 Pro model card) | Evaluation – Results (table, PDF p. 5) |
| Global PIQA | Claude Sonnet 4.5 | 90.1% | commonsense reasoning across 100 languages and cultures; reported by Google DeepMind in the Gemini 3 Pro model card; computed by Google DeepMind through the provider's official API | Evaluation – Results (table, PDF p. 5) |
| Global PIQA | GPT-5.1 | 90.9% | commonsense reasoning across 100 languages and cultures; reported by Google DeepMind in the Gemini 3 Pro model card; computed by Google DeepMind through the provider's official API | Evaluation – Results (table, PDF p. 5) |
| Text to Text Safety — Delta vs. Gemini 2.5 Pro | Gemini 3 Pro | -10.4% | absolute percentage change vs. Gemini 2.5 Pro on Google's internal automated evaluation (not human evaluation or red teaming); direction of 'better' is shown in the card only by colour, which the text capture does not preserve; developer-reported | Ethics and Content Safety – Training and Development Evaluation Results |
| Multilingual Safety — Delta vs. Gemini 2.5 Pro | Gemini 3 Pro | +0.2% | absolute percentage change vs. Gemini 2.5 Pro on Google's internal automated evaluation (not human evaluation or red teaming); direction of 'better' is shown in the card only by colour, which the text capture does not preserve; card marks the change '(non-egregious)'; developer-reported | Ethics and Content Safety – Training and Development Evaluation Results |
| Image to Text Safety — Delta vs. Gemini 2.5 Pro | Gemini 3 Pro | +3.1% | absolute percentage change vs. Gemini 2.5 Pro on Google's internal automated evaluation (not human evaluation or red teaming); direction of 'better' is shown in the card only by colour, which the text capture does not preserve; card marks the change '(non-egregious)'; developer-reported | Ethics and Content Safety – Training and Development Evaluation Results |
| Tone — Delta vs. Gemini 2.5 Pro | Gemini 3 Pro | +7.9% | absolute percentage change vs. Gemini 2.5 Pro on Google's internal automated evaluation (not human evaluation or red teaming); direction of 'better' is shown in the card only by colour, which the text capture does not preserve; positive = improvement in tone of refusals (card footnote 2); developer-reported | Ethics and Content Safety – Training and Development Evaluation Results |
| Unjustified-refusals — Delta vs. Gemini 2.5 Pro | Gemini 3 Pro | +3.7% | absolute percentage change vs. Gemini 2.5 Pro on Google's internal automated evaluation (not human evaluation or red teaming); direction of 'better' is shown in the card only by colour, which the text capture does not preserve; card marks the change '(non-egregious)'; developer-reported | Ethics and Content Safety – Training and Development Evaluation Results |
| Cybersecurity key skills benchmark — v1 hard challenges | Gemini 3 Pro | 11/12 | challenges solved; alert threshold met, Cyber Uplift Level 1 CCL not reached; developer-reported (detail in the Gemini 3 Pro Frontier Safety Framework report) | Frontier Safety |
| Cybersecurity key skills benchmark — v2 challenges | Gemini 3 Pro | 0/13 | challenges solved end-to-end; developer-reported | Frontier Safety |
| Situational awareness challenges (Phuong et al. 2025) | Gemini 3 Pro | 3/11 | challenges solved by the agent; Instrumental Reasoning Levels 1+2 CCL not reached; developer-reported | Frontier Safety |
| Stealth challenges (Phuong et al. 2025) | Gemini 3 Pro | 1/4 | challenges solved by the agent; developer-reported | Frontier Safety |
Provenance
The page summarises a text extraction of the PDF captured on September 19, 2026, with the results table transcribed from the page image, together with the methodology page. All figures are reported by Google DeepMind; comparison-model figures are a mix of the other developers' published numbers, third-party leaderboards and Google's own runs through provider APIs, as the methodology page specifies. Confidence is medium: a single developer-authored source, with fast-decay benchmark claims.