Gemini 2.5 Pro Model Card (Google DeepMind, updated June 2025)

The Gemini 2.5 Pro model card is a document published by Google DeepMind for Gemini 2.5 Pro; the version summarised here is marked "Model card updated June 27, 2025" (Source: https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-5-Pro-Model-Card.pdf). The card states that it was updated to cover Gemini 2.5 Pro GA in addition to Gemini 2.5 Pro Experimental (03-25) and Gemini 2.5 Pro Preview (05-06), and to reflect the model's move to general availability. It directs readers who want more detail to the Gemini 2.5 technical report. Because it reports the three versions side by side, the card records how results moved between the March, May and June 2025 releases.

Model information

The card describes Gemini 2.5 Pro as a natively multimodal reasoning ("thinking") model that accepts text, images, audio and video with a 1M-token context window and produces text with a 64K-token output. The Gemini 2.5 models are described as sparse mixture-of-experts transformers; parameter counts are not given. Training data is described as publicly available web documents, code, images, audio and video, with post-training on vetted instruction-tuning, human-preference and tool-use data; filtering included deduplication, safety filtering and quality filtering. Training used Google's Tensor Processing Units with JAX and ML Pathways. The knowledge cutoff is January 2025. Known limitations listed are hallucinations and limits in causal understanding, complex logical deduction and counterfactual reasoning; the main safety limitations named are over-refusals and tone.

Capability evaluation

The results table, labelled "as of June 2025", has columns for Gemini 2.5 Pro GA, Preview (05-06) and Experimental (03-25), and for OpenAI o3 and o4-mini at high reasoning, Claude Sonnet 4, Claude Opus 4 with 32K thinking, Grok 3 Beta with extended thinking, and DeepSeek R1 05-28. It covers Humanity's Last Exam, GPQA diamond, AIME 2025, LiveCodeBench (two leaderboard date windows), Aider Polyglot, SWE-bench Verified (single and multiple attempts), SimpleQA, FACTS Grounding, MMMU, Vibe-Eval, VideoMME, VideoMMMU, MRCR and MRCR v2 (8-needle), and Global MMLU (Lite). The card states that "the most recent model shows significant improvement in code performance" while other capabilities are similar across the three versions.

The approach section states that all Gemini results are pass@1 through the AI Studio API with default sampling, names the three model IDs used, and defines "single attempt" as allowing no majority voting or parallel test-time compute. It specifies which thinking settings were used for each comparison model, says that where a provider reported two numbers the higher used majority voting with n=64 (Grok) or parallel test-time compute (Anthropic), and states that SWE-bench Verified figures follow each provider's own scaffolding. For MRCR, the card says it changed methodology to a harder 8-needle version for the newest column, so the earlier versions' MRCR figures are not comparable with the GA figure.

Content safety

Automated safety evaluations are reported as absolute percentage changes relative to Gemini 1.5 Pro 002 for each of the three versions: text-to-text safety, multilingual safety, image-to-text safety, tone, and instruction following. The card also describes baseline assurance evaluations, human and automated red teaming, and mitigations including dataset filtering, supervised fine-tuning, reinforcement learning from human and critic feedback, and product-level safety filtering. No absolute violation rates are given.

Frontier safety

The card summarises evaluation against the critical capability levels (CCLs) of Google's Frontier Safety Framework, with separate columns for the three versions, and reports no CCL reached. For cybersecurity it gives key-skills benchmark solve counts for all three versions and autonomous-offense-suite counts for the Experimental version, and states that Gemini 2.5 Pro GA "does not show meaningful improvement" over the two earlier versions on the key-skills benchmark. For machine learning R&D it gives an average normalised RE-Bench score of 0.73 for the Experimental version and, for GA, best runs scoring between 50% and 125% of the best human-written solutions, below an alert threshold that the card says was set above human performance. Deceptive-alignment results (2 of 5 stealth and 2 of 11 situational-awareness challenges) are for the Experimental version; the card says those evaluations had not been completed on GA. CBRN results are qualitative, with multiple-choice benchmark scores shown only as a chart.

Differences from the technical report

Several figures in the card differ from those in the later technical report for what both call Gemini 2.5 Pro. The card gives 69.0% on LiveCodeBench (1/1/2025–5/1/2025 window) where the report gives 74.2%; key-skills counts of 7/8 easy, 10/28 medium and 1/12 hard for GA where the report, describing versions "up to 06-05", gives 7/8, 14/28 and 6/12; and a "Disabled Tool" success rate of 46% where the report gives 42%. Neither document explains the differences.

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.

Provenance

The page summarises the PDF's text layer as captured on September 19, 2026. Figures 1 to 4 are charts whose values are not given in the text. All results are reported by Google DeepMind; comparison figures are mostly the other developers' own numbers or public leaderboards. Confidence is medium.

Relationships

• supports: Gemini 2.5 Pro
• instance-of: AI System Cards
• related: Gemini 2.5 Technical Report (Google, July 2025), Gemini 3 Pro Model Card (Google DeepMind, November 2025), Google DeepMind, AI Safety Cases and Frameworks