Codex makes fewer bugs, but more people use Claude (www.cubic.dev)

A recent report on the state of AI coding reveals that while GPT-5.5 generates the cleanest code with the fewest bugs, Claude Code remains the dominant choice among developers, being used by 80% of them. Data sourced from cubic's review of thousands of code commits shows that GPT-5.5 produced fewer bugs per line of code than any competing models, including those in Anthropic's flagship offerings. However, more developers prefer Claude Code, with the Opus model family leading as the primary coding agent in 80% of cases. This trend underscores a significant divergence between model quality and user preference. In addition to model comparisons, the report highlights a remarkable surge in developer productivity, with the number of merged pull requests (PRs) tripling over the past year and the median size of these PRs increasing by 2.2 times. Most code is now authored by AI, accounting for 90-100% among teams using cubic's tracking tool, marking a shift towards a more AI-centric coding environment. The swift adoption of new models, like Claude Fable—despite its brief availability—further illustrates the rapid evolution and competitive dynamics within the AI/ML landscape, suggesting that performance metrics alone may not dictate model preference among developers.