Google has released Gemini 2.5 Pro — an experimental version of the most advanced model that mentally articulates its answer before delivering it. This is not just generation, but full-fledged reasoning: analysis, conclusions, context, logic.
As a result, the model has reached the top position in LMArena (a rating based on human preferences) and has outperformed competitors in mathematical, scientific, and coding benchmarks.
What’s particularly interesting:
- 18.8% on Humanity’s Last Exam — an exam on frontier skills devised by hundreds of experts;
- 63.8% on SWE-Bench Verified — one of the main tests for “agent programming”;
- context window — up to a million tokens already now, two million — coming soon.
Gemini 2.5 can generate complex web applications, edit and transform code, and even create games from a single prompt. Moreover, it does this based on reasoning, rather than just guessing tokens.
It will soon be available on Vertex AI. For now, you can play in Google AI Studio or in Gemini Advanced.
