August 2025 Summaries

Cursor has introduced an integration with Linear, enabling users to trigger Cursor Background Agents directly from the Linear platform to address bugs, build features, and respond to user feedback efficiently. This integration allows users to delegate tasks to Cursor agents by assigning issues or mentioning Cursor in comments, facilitating automatic triaging and seamless collaboration among team members. Cursor pulls in contextual information from issues, comments, and references to ensure agents have the necessary details to complete tasks. Notifications are sent once agents complete tasks and generate pull requests, allowing teams to review and manage coding tasks without leaving their project management workflow. The Cursor team itself utilizes this integration to manage thousands of tasks, including product bugs and customer feedback, by automatically assigning these to Cursor agents for prompt resolution and enhancement of their repository.

The blog post discusses a significant optimization achieved in training large language models on Blackwell GPUs (B200s) using custom MXFP8 kernels, resulting in a 1.5x overall training speedup compared to the previous Hopper GPU (H100) setup. By identifying the Mixture-of-Experts (MoE) layer as a key bottleneck, the team rewrote it from scratch at the GPU kernel level with minimal dependencies, leveraging pure CUDA and PTX. They introduced microscaling with MXFP8 data formats to reduce computational costs, achieving a 3.5x improvement in MoE layer performance with almost no loss in training quality. The transition to block-scaled matrix multiplication and the development of a highly optimized MXFP8 quantization kernel were crucial in maximizing the performance on Blackwell GPUs, overcoming challenges related to new tensor memory architecture and quantization overhead. This comprehensive enhancement allows their stack to outperform open-source alternatives, paving the way for faster MoE training and future exploration of FP4 precision.

The recent updates announced for the billing model of Teams and Auto reflect a shift towards a more flexible and competitive pricing strategy, set to take effect from the next billing renewal after September 15 for both monthly and yearly plans. The Teams plan will transition from fixed to variable request costs, allowing charges to fluctuate based on the complexity of the agent's tasks. This aligns the Teams billing system with the individual plans, which already use a similar credit system. Additionally, the pricing and usage limits for Auto are being adjusted to contribute towards monthly usage at competitive token rates. Previously, Auto was priced similarly to other premium models and was unlimited for individuals from June 2025. These changes aim to adapt to an agentic world and enhance user experience, with a commitment to improved quality and performance.

GPT-5, OpenAI's most powerful model, is now integrated into Cursor and has shown remarkable effectiveness in coding tasks according to initial feedback from engineers. The model has impressed users with its ability to handle long-running and complex tasks, such as resolving intricate bugs and optimizing Stripe queries to reduce latency. Engineers have found GPT-5 to be highly steerable, requiring explicit instructions to perform tasks accurately, and it has demonstrated the ability to regenerate code components like protobufs across frontend and backend systems. Despite its default verbose output, users have successfully adjusted the model's behavior to suit their preferences for conciseness and agentic responses. As a launch offer, GPT-5 is accessible to paying users with free credits during the first week of its release.

Cursor has introduced the capability to use its Cursor Agent from the Command Line Interface (CLI) or in a headless mode across various environments, enhancing accessibility for developers who prefer different Integrated Development Environments (IDEs) like Neovim and JetBrains. Initially designed to improve code autocompletion and facilitate AI-driven code enhancements, Cursor Agent can now generate files, run terminal commands, and search codebases across platforms such as web, mobile, and Slack. This expansion allows developers to leverage AI capabilities in their preferred tools, with the CLI working with any model under a Cursor subscription, enabling agents to operate in parallel within terminals or remotely. Although still in beta, the CLI offers functionalities like file modification and shell command execution, albeit with evolving security measures, and users are encouraged to provide feedback as they experiment with these new features.