The Gemini 3.1 Pro model card is a document published by Google DeepMind on February 19, 2026, alongside the release of Gemini 3.1 Pro, a point-update in the Gemini 3 series (Source: https://deepmind.google/models/model-cards/gemini-3-1-pro/). The card describes the model as "based on Gemini 3 Pro" and refers the reader to the Gemini 3 Pro model card for architecture, training data, hardware, software, known limitations, acceptable use, safety policies, and risks and mitigations. Its own content is a benchmark table, a table of automated safety-evaluation changes relative to Gemini 3 Pro, a statement on human red teaming, and a Frontier Safety Framework assessment. A companion page gives the evaluation methodology (Source: https://deepmind.google/models/evals-methodology/gemini-3-1-pro/).
Model information and distribution
The card describes Gemini 3.1 Pro 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. It lists seven distribution channels: the Gemini app, Google Cloud / Vertex AI, Google AI Studio, the Gemini API, Google Antigravity, Gemini Enterprise and NotebookLM. Google states the model is suited to agentic performance, advanced coding, long-context and multimodal understanding, and algorithmic development.
Capability evaluation
The results table, labelled "as of February 2026", compares Gemini 3.1 Pro at Thinking (High) with Gemini 3 Pro at Thinking (High), Claude Sonnet 4.6 and Claude Opus 4.6 at Thinking (Max), and GPT-5.2 and GPT-5.3-Codex at Thinking (xhigh). It covers Humanity's Last Exam (without tools, and with search and code), ARC-AGI-2, GPQA Diamond, Terminal-Bench 2.0, SWE-Bench Verified, SWE-Bench Pro (Public), LiveCodeBench Pro, SciCode, APEX-Agents, GDPval-AA, τ2-bench (Retail and Telecom), MCP Atlas, BrowseComp, MMMU-Pro, MMMLU and MRCR v2 (8-needle) at 128k and 1M tokens. The card states that Gemini 3.1 Pro "significantly outperforms Gemini 3 Pro across a range of benchmarks requiring enhanced reasoning and multimodal capabilities". In the table, Gemini 3.1 Pro has the highest figure in most rows; Claude Opus 4.6 is highest on Humanity's Last Exam with search and code, SWE-Bench Verified and τ2-bench Retail, Claude Sonnet 4.6 on GDPval-AA, GPT-5.3-Codex on SWE-Bench Pro and on Terminal-Bench 2.0 under its own harness, and Gemini 3 Pro on MMMU-Pro.
The methodology page states that all Gemini scores are pass@1 unless noted, that "single attempt" settings allow no majority voting or parallel test-time compute, that runs used the Gemini API model gemini-3.1-pro-preview with default sampling, and that results for non-Gemini models are the providers' self-reported numbers at their maximum reasoning settings where available. Several figures are taken from third parties: ARC-AGI-2 from the ARC Prize website (ARC Prize Verified, semi-private set), LiveCodeBench Pro and Terminal-Bench 2.0 comparison figures from public leaderboards, SciCode and GDPval-AA from Artificial Analysis, and MCP Atlas from the public set. For SWE-Bench Verified, Google reports that it found three items that cannot pass on the official harness, that Gemini 3.1 Pro passed all three under a corrected internal implementation, and that it raised the model's score by 0.6% to reflect this. The τ2-bench Airline domain was excluded "due to lower quality grading". Google states it is releasing the MRCR v2 dataset for reproducibility.
Content safety
The card reports five internal automated safety evaluations as absolute percentage changes relative to Gemini 3 Pro: text-to-text safety, multilingual safety, image-to-text safety, tone of refusals, and unjustified refusals. All five changes are smaller than half a percentage point. Google states that Gemini 3.1 Pro "outperforms Gemini 3 Pro across both safety and tone, while keeping unjustified refusals low", that results are not comparable with earlier model cards because the evaluations were revised, and that manual review found losses were "overwhelmingly either a) false positives or b) not egregious". On human red teaming, the card says Gemini 3.1 Pro satisfied required launch thresholds for child safety and showed safety performance similar to Gemini 3 Pro across content policies. No absolute violation or refusal rates are given.
Frontier safety
Google reports a full Frontier Safety Framework evaluation of Gemini 3.1 Pro, run mainly in Deep Think mode, and concludes that no critical capability level (CCL) was reached in any of five domains.
- CBRN. The model "can provide highly accurate and actionable information" but does not give novel or sufficiently complete instructions for critical stages; the alert threshold was not reached. Mitigations remain deployed.
- Cyber. Because Gemini 3 Pro had reached the alert threshold, Google ran additional testing with and without Deep Think. The model shows increased cyber capability over Gemini 3 Pro and has reached the alert threshold but not the CCL. With Deep Think, accounting for inference cost, it performs "considerably worse" than without.
- Harmful manipulation. Maximum odds ratio of 3.6x on belief-change metrics against a non-AI baseline, the same as Gemini 3 Pro; alert threshold not reached.
- Machine learning R&D. A human-normalised average of 1.27 on RE-Bench against 1.04 for Gemini 3 Pro. On the Optimise LLM Foundry task the model cut a fine-tuning script's runtime from 300 seconds to 47 seconds, against 94 seconds for the human reference solution. Average performance remains below the alert threshold.
- Misalignment (exploratory). Stealth results are similar to Gemini 3 Pro. On situational awareness the model reaches a success rate of "almost 100%" on three challenges (max tokens, context size mod, oversight frequency) that the card says no other model had consistently solved, but is inconsistent on the others.
The card refers to the Gemini 3 Pro Frontier Safety Framework Report for evaluation and mitigation detail; no separate report for Gemini 3.1 Pro is cited.
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 |
|---|---|---|---|---|
| SciCode | Gemini 3.1 Pro | 59% | Thinking (High); sourced from Artificial Analysis; developer-reported | Evaluation – Results |
| SciCode | Gemini 3 Pro | 56% | Thinking (High); sourced from Artificial Analysis; developer-reported (Google DeepMind, in the Gemini 3.1 Pro model card) | Evaluation – Results |
| SciCode | Claude Sonnet 4.6 | 47% | Thinking (Max); sourced from Artificial Analysis; reported by Google DeepMind in the Gemini 3.1 Pro model card; sourced from the provider's self-reported numbers unless the card's methodology page says otherwise | Evaluation – Results |
| SciCode | Claude Opus 4.6 | 52% | Thinking (Max); sourced from Artificial Analysis; reported by Google DeepMind in the Gemini 3.1 Pro model card; sourced from the provider's self-reported numbers unless the card's methodology page says otherwise | Evaluation – Results |
| SciCode | GPT-5.2 Thinking | 52% | Thinking (xhigh); sourced from Artificial Analysis; reported by Google DeepMind in the Gemini 3.1 Pro model card; sourced from the provider's self-reported numbers unless the card's methodology page says otherwise | Evaluation – Results |
| Text to Text Safety — Delta vs. Gemini 3 Pro | Gemini 3.1 Pro | +0.10% | absolute percentage change vs. Gemini 3 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 |
| Multilingual Safety — Delta vs. Gemini 3 Pro | Gemini 3.1 Pro | +0.11% | absolute percentage change vs. Gemini 3 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 3 Pro | Gemini 3.1 Pro | -0.33% | absolute percentage change vs. Gemini 3 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 |
| Tone — Delta vs. Gemini 3 Pro | Gemini 3.1 Pro | +0.02% | absolute percentage change vs. Gemini 3 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; developer-reported | Ethics and Content Safety – Training and Development Evaluation Results |
| Unjustified-refusals — Delta vs. Gemini 3 Pro | Gemini 3.1 Pro | -0.08% | absolute percentage change vs. Gemini 3 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 |
| Harmful manipulation efficacy (belief change) — Maximum odds ratio vs. non-AI baseline | Gemini 3.1 Pro (Deep Think mode) | 3.6x | maximum odds ratio on belief-change metrics against a non-AI baseline; card says this equals Gemini 3 Pro and is below the alert threshold; developer-reported | Frontier Safety |
| Harmful manipulation efficacy (belief change) — Maximum odds ratio vs. non-AI baseline | Gemini 3 Pro | 3.6x | stated in the Gemini 3.1 Pro card as 'the same as Gemini 3 Pro'; developer-reported (Google DeepMind, in the Gemini 3.1 Pro model card) | Frontier Safety |
Provenance
The page summarises the HTML card as captured on September 19, 2026, together with its methodology page. The card links a PDF version (Source: https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-1-Pro-Model-Card.pdf). All benchmark and safety figures are reported by Google DeepMind; comparison-model figures are mostly the other developers' own published numbers, selected by Google. Confidence is medium: a single developer-authored source, with fast-decay benchmark claims.