Gemini 2.5 Pro Experimental
2025-03-25 · Publisher : Google · Domain : Artificial Intelligence
Google launches Gemini 2.5 Pro Experimental, the first model of the 2.5 family: a reasoning model ranked first on LMArena, strong in code, mathematics and science, with 1 million tokens of context.
What changes
- “Thinking” model: it reasons before answering, a capability now built into all future Gemini models.
- First on LMArena by a significant margin, and state of the art in mathematics and science (GPQA, AIME 2025) without costly majority-voting techniques.
- Big progress in code compared with the 2.0 generation: web applications, agents, code transformation and editing.
- Context of 1 million tokens (2 million announced), with text, audio, image, video and code repository inputs.
### Specifications - Context: 1,048,576 input tokens, 65,536 output tokens (current API documentation). - Knowledge cutoff: January 2025 (API documentation). - Availability: Google AI Studio and the Gemini app (Advanced subscribers); Vertex AI afterwards. - Identifier: `gemini-2.5-pro-exp-03-25`. Paid version `gemini-2.5-pro-preview-03-25` on April 4, 2025. ### Pricing | Version | Input (USD / 1M tokens) | Output (USD / 1M tokens) | |---|---|---| | Paid preview of April 4, prompts up to 200,000 tokens | 1.25 | 10 | | Paid preview, prompts of more than 200,000 tokens | 2.50 | 15 | Free experimental version with reduced rate limits. The paid preview pricing was noted by Simon Willison on April 4, and is identical to that of the stable model. ### Benchmarks | Benchmark | Score | |---|---| | Humanity's Last Exam (no tools) | 18.8% | | SWE-bench Verified (custom agent) | 63.8% | Scores published by the vendor. The GPQA and AIME 2025 results appear in the post as a chart. ### Key takeaways The model that puts Google back at the top of the reasoning rankings, the basis of the Pro range used in production throughout 2025.
Links
Tags : LLM, Google, Gemini, Pro, Raisonnement