Gemini 2.5 Pro (Experimental)
Gemini 2.5 Pro is the first 2.5 model, a thinking model that debuted #1 on LMArena by a wide margin, with 18.8% on Humanity's Last Exam.
Google combined a stronger base model with improved post-training and said thinking would be built into all future Gemini models. 1M-token context at launch (2M promised). Coding is a big leap over 2.0.
- Date
- Tuesday, 25 March 2025
- Lab
- Google DeepMind
- Kind
- model
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| Humanity's Last Exam (no tools) | 18.8% state of the art across models at launch | company |
| SWE-Bench Verified | 63.8% custom agent setup | company |
| LMArena | #1 by a significant margin at debut | LMArena |
Experimental on 2025-03-25; billing-enabled preview 2025-04-04; stable GA 2025-06-17. An updated "I/O edition" reached developers early, about two weeks before I/O, per the I/O 2025 Gemini post.
Sources
- blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/
- ai.google.dev/gemini-api/docs/changelog
This record was checked against its sources on 6 October 2026. How we check