We Made Haiku as Good as Opus. Improving Claude Code with Codeset (codeset.ai)

Codeset has launched a powerful tool that enhances the coding agent Claude Code by improving its task resolution rates across its various models. Specifically, Claude Haiku 4.5's success rate increased from 52% to 62%, Sonnet 4.5 jumped from 56% to 65.3%, and Opus 4.5 went from 60.7% to 68% on a set of real-world engineering tasks. By leveraging historical project data from GitHub repositories, Codeset provides valuable insights such as past bugs, testing requirements, and co-dependencies among files, giving coding agents the contextual understanding necessary to avoid common pitfalls and optimize performance. This innovation is significant for the AI/ML community as it demonstrates that structured context can enhance even lower-tier models, allowing teams to achieve better results at a fraction of the operational cost. For example, Haiku with Codeset can now outperform raw Opus while being ten times cheaper per task. Moreover, the validation across two distinct benchmark datasets, codeset-gym-python and SWE-Bench Pro, affirms that effective context extraction can sustain improvements across various programming languages and typical software engineering challenges. This positions Codeset as a transformative solution for development teams looking to maximize productivity and minimize costs in software development using AI-driven agents.