Gemini 2.5 Pro is a frontier model from Google DeepMind, released in experimental form on March 25, 2025 and reaching general availability in June 2025. It sits between Google's long-context Gemini 1.5 / 2.0 generation and the 2026 flagship Gemini 3 / Gemini 3 Pro. It introduced reasoning enabled by default, a native 1M-token context window, and held top positions on most public benchmarks for roughly March through August 2025. (Source: State of AI Report 2025) (Source: epochai.substack.com)
| Field | Value | |
|---|---|---|
| Developer | [[google-deepmind | Google DeepMind]] |
| Released | March 25, 2025 (experimental); general availability June 2025 | |
| Architecture | Sparse MoE (details undisclosed) | |
| Context | 1M tokens native, 2M in experimental preview | |
| Modalities | Text, image, audio, video input; text + image + audio output | |
| Distribution | Gemini app, Vertex AI, Google AI Studio | |
| License | Closed |
Model family
The Gemini 2.5 line spans several tiers:
- Gemini 2.5 Pro — flagship, reasoning-enabled by default.
- Gemini 2.5 Flash — fast tier with optional thinking mode.
- Gemini 2.5 Flash-Lite — cheapest tier, released later.
- Gemini 2.5 Deep Think — June 2025; extended parallel-sampling reasoning mode.
Capabilities and benchmarks
Gemini 2.5 Pro was the first Google model to match or exceed the GPT-4 generation and Claude rivals across both reasoning and multimodal axes, according to (Source: State of AI Report 2025) and (Source: epochai.substack.com). At release it held #1 positions on most public benchmarks, including LMSYS Chatbot Arena, GPQA, AIME, and Humanity's Last Exam. For roughly March through August 2025 it was the most-cited top frontier model in evaluation papers, before GPT-5.4 Thinking, Claude Opus 4.6, and Gemini 3 / Gemini 3 Pro re-ranked the field.
Two features set it apart from prior Gemini releases. First, native reasoning: the flagship always "thinks" before responding, with no non-thinking mode (Flash exposes thinking as optional), paralleling OpenAI o-series (o1 → o4-mini) and Claude's extended-thinking mode. Second, the 1M-token native context is processed in a single forward pass rather than through retrieval augmentation; effective long-context performance on needle-in-haystack tests exceeded 99% at full length.
Benchmark results at release, compared with o3 and Claude 3.7 Sonnet:
| Benchmark | Gemini 2.5 Pro | o3 | Claude 3.7 Sonnet |
|---|---|---|---|
| GPQA-Diamond | 84.0 | 87.7 | 78.2 |
| AIME 2025 | 86.7 | 88.9 | 61.3 |
| Humanity's Last Exam | 18.8 | 20.3 | 8.9 |
| SWE-Bench Verified | 63.8 | 69.1 | 62.3 |
| MMMU (multimodal) | 81.7 | — | 75.0 |
| LMSYS Arena | #1 (Mar–May 2025) | — | — |
Its benchmark standing prompted several benchmark-saturation discussions that contributed to Humanity's Last Exam, ARC-AGI-2, and FrontierMath becoming primary reporting axes. The model's emergence is also cited as marking Google's return to frontier parity after a period in which the comparison between Anthropic and OpenAI dominated frontier discussion, in the context of US-China AI competition. The same trajectory is treated in (Source: epochai.substack.com) and (Source: oneusefulthing.org) as the 2025 midpoint between GPT-4 and Gemini 3.
Training and architecture
Gemini 2.5 Pro uses a sparse mixture-of-experts (MoE) architecture; total and active parameters are undisclosed. Multimodal generation includes native image and audio output, closing a gap that prior Gemini versions had versus GPT-4o.
The Gemini 2.5 Deep Think mode, released in June 2025, spends substantially more inference compute on a single prompt through extended parallel sampling, raising AIME to roughly 99% and GPQA to roughly 87%.
Safety and evaluations
Google DeepMind ran Frontier Safety Framework evaluations — the DeepMind analogue of Anthropic's Responsible Scaling Policy (Version 3.1) and OpenAI Preparedness Framework V.2 — on Gemini 2.5 Pro prior to release. Public results reported CBRN uplift below concern thresholds, cyber-capability evaluations on par with prior Gemini generations, and persuasion evaluations that showed the first DeepMind model with measurable persuasion uplift over the prior generation.
Gemini 2.5 Pro is part of the "covered frontier" under Frontier AI Safety Commitments (Seoul, 2024), with published reporting for both the UK and US AI Safety Institutes. (Source: UK AISI — Advanced AI Evaluations May Update)
Availability and pricing
The model is distributed through the Gemini app, Vertex AI, and Google AI Studio under a closed license. It supports 1M tokens of native context, with a 2M-token configuration offered in experimental preview.
Relationships
- superseded-by: Gemini 3 / Gemini 3 Pro
- instance-of: General-Purpose AI (GPAI)
- related: Google DeepMind, Demis Hassabis, OpenAI o-series (o1 → o4-mini), Claude Opus 4.6, GPT-4 Family (OpenAI), (Source: epochai.substack.com), Frontier AI Safety Commitments (Seoul, 2024)