Aider has introduced an innovative approach to coding tasks by separating the processes of code reasoning and code editing. This new method employs two distinct models: an Architect model, which is responsible for conceptualizing how to solve a coding problem, and an Editor model, which takes the Architect's solution and generates specific code editing instructions to modify existing source files. This separation has led to state-of-the-art (SOTA) results in Aider's code editing benchmarks. The motivation behind this dual-model approach stems from the capabilities of OpenAI's o1 models, which excel in reasoning but often struggle to produce well-formatted code editing instructions. By allowing the Architect to describe the solution in its own terms and then passing that output to a more traditional language model (LLM) for editing, Aider can enhance the quality of code modifications. This method has gained traction due to recent advancements in the speed and cost-effectiveness of cutting-edge models, making it feasible to implement in an interactive coding environment. Traditionally, Aider would prompt a single model to both solve a coding problem and return formatted editing instructions in one go. This required the model to juggle the complexities of problem-solving and adhering to specific formatting requirements simultaneously. The new Architect/Editor framework divides this task into two steps: first, the Architect focuses solely on solving the problem, and then the Editor concentrates on formatting the necessary code edits. This division allows each model to operate within its strengths, with the Architect handling reasoning and the Editor managing the technicalities of code formatting. The results of this new approach have been promising. Benchmark scores indicate that pairing the o1-preview model as the Architect with either DeepSeek or o1-mini as the Editor achieved a remarkable score of 85%. This configuration, while effective, is noted to be slower and may not be practical for real-time interactions. However, using o1-preview with Anthropic's Sonnet as the Editor also yielded strong results, demonstrating a viable option for users who can access both models. The performance improvements are not limited to specific model pairings; many models benefit from the Architect/Editor configuration. For instance, models like Sonnet, GPT-4o, and GPT-4o-mini showed enhanced scores when used in this dual role setup. DeepSeek has also proven to be an effective Editor, adept at transforming proposed solutions into updated source files, particularly when using the efficient "diff" editing format. To facilitate this new functionality, Aider's development version includes built-in defaults for Architect/Editor coding with various models. Users can easily set up their environment to utilize these capabilities by installing Aider and configuring their API keys for the desired models. The introduction of the "architect" chat mode further streamlines the process, allowing users to engage with the system more intuitively. Overall, the separation of code reasoning and editing in Aider represents a significant advancement in AI-assisted programming, enhancing both the efficiency and effectiveness of coding tasks. The results from various model combinations underscore the potential of this approach, paving the way for more sophisticated and user-friendly coding experiences.