Gemini 2.5 Pro – a new top among language models
Google has released Gemini 2.5 Pro Preview 06-05, an updated version of its most powerful LLM, which is also one of the most powerful neural networks on the market. The company claims leadership in most of the known rankings, with the model competing with the past version of itself in some.
The model improved markedly in AIDER Polyglot (a benchmark to assess multilingual programming ability), HLE (a test of deep reasoning and general knowledge) and GPQA (a benchmark of graduate/postgraduate level physics questions testing the model’s scientific and mathematical abilities).
The model scores 1,443 points in WebDev Arena, a ranking that shows the success of neural networks in web development.
1,470 points in the Leaderbord Text Arena, a ranking in which models compete in fiction writing.
1278 points in Vision Arena, a benchmark that evaluates multimodality:
Google also claims to have worked on the model’s communication style – now her answers should become more witty and creative, as well as better structured. But this statement everyone should check for themselves.
The model has already appeared in Google AI Studio, and in the near future it will surely become available in the Gemini app for Pro and Ultra subscribers. It’s not without a little negativity here: just two days ago, Google introduced limits on the use of 2.5 Pro for 20-dollar subscription users, limiting them to 50 requests per day. After a mountain of outraged X posts, the limits were increased to 100 queries per day. That’s less, since 2.5 Pro used to be virtually unlimited, but it’s still better than the ChatGPT Plus subscription, where the competing o3 is limited to 100 requests per week.
