Summary

  • Google has launched Gemini 2.5 Pro Experimental, its “most intelligent AI model” yet, which has already topped the LMArena leaderboard.
  • A key feature of Gemini 2.5 Pro is its enhanced ability to “think and reason” before responding, leading to improved performance and accuracy in complex tasks.
  • Currently available for Gemini Advanced users on the web (mobile support often lags behind), 2.5 Pro outperformed competitors in reasoning, knowledge, science, and math benchmarks, and has a 1 million-token context window.

Google is constantly working on enhancing Gemini’s capabilities. Gemini 2.0 Advanced landed on mobile earlier this year, followed by enhancements to the AI tool within Workspace apps. Subsequently, the tech giant expanded Gemini 2.0 Flash access to all, followed by the rollout of its former top-of-the-line model Gemini 2.0 Pro Experimental, with Gemini 2.0 Flash Thinking Experimental and Flash Thinking Experimental with apps in tow.

You’d be wondering why I described 2.0 Pro as Gemini’s former top-of-the-line model. This is because Google today began rolling out Gemini 2.5 Pro, which the tech giant describes as its “most intelligent AI model.”

two phones running google gemini app and gemini live on a laptop keyboard


Related


Google Gemini: Everything you need to know about Google’s next-gen multimodal AI

Google Gemini is here, with a whole new approach to multimodal AI

From the new 2.5 family, Google has only begun rolling out Gemini 2.5 Pro Experimental, a model that is reportedly designed to tackle increasingly complex problems. It debuted at the #1 spot on the community-driven LMArena LLM leaderboard‘s ‘overall’ category, followed by the Grok 3 Beta.

A key highlight of the new model is its ability to think and reason with its own thoughts before responding — a quality that us as humans could benefit from. The tech giant suggests that this results in enhanced performance and improved accuracy, likely aiding in avoiding hallucinations.”Going forward, we’re building these thinking capabilities directly into all of our models, so they can handle more complex problems and support even more capable, context-aware agents,” Google indicated.

The new model was compared with others in its league, including OpenAI’s o3-mini and GPT-4.5, Claude’s Sonnet 3.7, Grok 3 Beta, and DeepSeek R1. Gemini 2.5 Pro managed to outperform all mentioned models when it comes to reasoning and knowledge, science-related queries, mathematics, code editing, visual reasoning, long-context reasoning, and more. It, however, did lag behind some models when it comes to code generation, agentic coding, and even factuality.

Notably, 2.5 Pro scored higher than all other comparable models in ‘Humanity’s Last Exam (no tools),’ a language model academic benchmark meant to test human knowledge of a wide range of subjects.

Humanity’s Last Exam: a dataset designed by hundreds of subject matter experts to capture the human frontier of knowledge and reasoning.

You can try out the new model now, but only if you have a Gemini Advanced subscription

Gemini 2.5 Pro Experimental has begun rolling out now for Gemini Advanced users. The model is available to us on the web but not on the mobile app — but that isn’t surprising. Support on mobile often lags by a few weeks.

The new model is currently limited to a 1 million token context window, with an upgrade to “2 million coming soon,” according to the tech giant.

A screenshot highlighting suport for Gemini 2.5 Pro Experimental.