There are all manner of coding benchmarks that rank the top AI models on the basis of the performance, but there might be a different way of comparing different models — by how much they’re being used by actual coders.
Cursor has come out with a list of the top AI models being used on its platform. Cursor is an AI coding IDE, that helps developers use AI to write code. It lets users choose between different models, so the choices made by developers seem like a pretty reliable way to rank AI models. Cursor is the most popular AI IDE, and its founder says that it now generates a billion lines of accepted code a day, so its sample size of coders is quite large too.
Cursor says that the most popular model on its platform in April 2025 was Anthropic’s Claude 3.7 Sonnet. Claude Sonnet has developed a reputation for being good at code, and this seems to be reflecting in the usage rankings. At second place is Google Gemini 2.5 Pro, the recently-released model from Google that’s topped many benchmarks. At third place is Claude 3.5 Sonnet. The fourth and fifth spots are occupied by two models from OpenAI, GPT 4.1 and 4o respectively.
Cursor also revealed the fastest growing models on its platform. Two OpenAI models topped this list, with the first two spots being occupied by the recently-released o3 and o4-mini. In third place was Chinese DeepSeek V3.1, while GPT-4.1 was in fourth place. The fifth spot was occupied by Elon Musk’s Grok 3.
And it’s this proliferation of models — of roughly equal quality — that’s put leverage in the hands of user-facing products like Cursor. Cursor owns the overall user-experience, while users have a choice to choose from one of many models available. This makes Cursor’s business model robust — moving code between different IDEs can be tricky — but users can seamlessly change the models that generate the code. It’s perhaps for this reason that Cursor was being courted by OpenAI for an acquisition, and reportedly has a ARR of $300 million. And if Cursor and other AI IDEs like Windsurf can keep releasing the top AI models on their platforms, they can give a great sense of popularity of different models based on their actual use.