The AI coding landscape has undergone significant changes recently, with major companies like Augment Code and warp.dev adjusting their pricing models due to sustainability concerns. Many firms are moving away from fixed base pricing because they do not own the inference infrastructure, which has shifted the competitive dynamics. Cursor’s release of its own model, Composer 1, marks a notable departure from reliance on external providers like Anthropic and OpenAI. This move, along with Windsurf’s acquisition by the Cognition team and their release of the fast Sweet 1.5 model running on Cerebras hardware, signals a new era where owning both the model and infrastructure is becoming crucial for success.
Composer 1, believed to be an upgraded version of the earlier Cheetah model, offers impressive speed at 250 tokens per second and is gaining traction among developers. The speaker highlights that many engineers do not prioritize the specific model used as long as it performs well, and both Composer 1 and Sweet 1.5 meet that standard. While these models may not be the absolute best in the market, they are sufficiently capable for about 75% of users, especially given their speed and cost-effectiveness. The shift towards owning proprietary models allows companies like Cursor to better control pricing and service quality, reducing dependency on external AI providers.
The speaker shares practical experiences using both Composer 1 and Sweet 1.5 across various projects, including a text-based adventure game, a Calendly clone, and a personal finance app. Composer 1 consistently demonstrated faster response times and more robust functionality, such as backend server setup and pre-trained data recognition, while Sweet 1.5 showed solid but simpler implementations. Despite some minor bugs and limitations, both models performed well enough to be integrated into real coding workflows. The speed advantage of these models enables real-time collaboration and rapid prototyping, which is highly valuable in professional environments.
Cursor 2.0 introduces new features like multi-model agents and work tree management, allowing users to run multiple models in parallel and compare outputs efficiently. This flexibility is particularly appealing for complex tasks, although it comes with higher costs. The interface improvements and chat-centric design enhance usability, making Cursor a compelling tool for developers who want to leverage AI coding assistants effectively. The speaker appreciates Cursor’s approach to enabling simultaneous model runs and managing workspaces, which many other AI coding tools have yet to address adequately.
In conclusion, the speaker believes Cursor and Windsurf are well-positioned for the future by developing and owning their AI models, despite some community concerns about the use of Chinese base models. While OpenAI and Anthropic models remain superior in some respects, the gap is narrowing, and the speed and cost benefits of Composer 1 and Sweet 1.5 make them attractive options. The evolving pricing models and infrastructure ownership will likely determine which companies thrive in the AI coding space. The speaker invites feedback from the community on these models and expresses optimism about the direction Cursor is heading, emphasizing that the AI coding game has fundamentally changed.
