Google releases experimental Gemini 2.5 Pro model

Google has released Gemini 2.5 Pro, an experimental version of its newest flagship AI model.

This is the first release in the Gemini 2.5 family and is now available to developers via Google AI Studio and to Gemini Advanced subscribers. Support for Vertex AI will follow soon.

The model is designed for complex reasoning tasks and long-context processing. Google says Gemini 2.5 Pro introduces a new system architecture that separates internal reasoning from response generation — allowing the model to process more information before producing an output. This approach improves accuracy in areas like logic, problem-solving, and decision-making.

Gemini 2.5 Pro “debuts at #1 on LMArena by a significant margin,” outperforming GPT-4 and Claude on a leaderboard based on human preferences. It also leads in key academic and technical evaluations:

• AIME 2025 for mathematics
• GPQA Diamond for science
• SWE-bench Verified, where it solves 63.8% of GitHub issues with a custom agent

In addition to reasoning improvements, the model expands its coding capabilities. Developers can build and edit full-stack web apps directly from prompts, with support for live previews, debugging, and code transformation.

Compared to Gemini 2.0 Flash, released in December, this version brings a notable performance jump — particularly in how it handles reasoning and code-heavy workflows. Gemini 2.5 Pro also achieves the top score (18.8%) on Humanity’s Last Exam, a benchmark designed to test models on the frontier of human knowledge.

Pricing and production-level usage limits will be introduced in the coming weeks. More models in the Gemini 1.5 and 2.5 lines are expected later this year.