Claude Sonnet 4.5: Coding, Agents, and Long-Context Evaluation

This episode primarily discusses the evaluation and performance of large language models (LLMs) in complex software engineering tasks, specifically focusing on long-context capabilities. One source, an excerpt from Simon Willison’s Weblog, praises the new Claude Sonnet 4.5 model for its superior performance in code generation, detailing an impressive complex SQLite database refactoring task it successfully completed using its Code Interpreter feature. The second source, an abstract and excerpts from the LoCoBench academic paper, introduces a new, comprehensive benchmark designed to test long-context LLMs up to 1 million tokens across eight specialized software development task categories and 10 programming languages, arguing that existing benchmarks are inadequate for realistic, large-scale code systems. This paper reveals that while models like Gemini-2.5-Pro may lead overall, different models, such as GPT-5, show specialized strengths in areas like Architectural Understanding. Finally, a Reddit post further contributes to the practical discussion by sharing real-world testing results comparing Claude Sonnet 4 and Gemini 2.5 Pro on a large Rust codebase.