Claude

Jason Clinton, Anthropic's Deputy CISO, wrote in July 2026 that Claude produces around 80% of the code merged at Anthropic, and that the internal Claude Tag instance merges more than half of it. David Thomas, adding features to a small graphic animation project with Claude, described the other end of the same speed: in 18 hours spread over five evenings he reached the kind of unmaintainable code teams normally take 18 months to accumulate. He calls the models "puppy-dog junior developers", eager to please, quick to suggest another feature, implicitly discouraging refactoring.

What recurs across the sources is Claude treated as one component inside a larger assembly. Katelyn Lesse, who leads engineering for Anthropic's developer platform, describes a configurable token budget for thinking API reasoning, MCP for external systems, a memory tool, context editing worth 39% on internal benchmarks, a sandboxed code execution environment, and Agent Skills for domain expertise. Thomas Kurian presented a Gemini agent that picks per task between Gemini models and Anthropic's Claude models. Guillaume Dumortier built a marketing system inside Claude where six of seven agents do nothing but check.

Rankings stay contested. SFEIR reads Claude as roughly 15 points ahead on SWE-Bench Pro while GPT-5.6 Sol leads Terminal-Bench 2.1, and argues Kimi K3 adds a portfolio column rather than a replacement. Felix Rieseberg, who works on Claude at Anthropic, locates the open problem elsewhere: capabilities advance unevenly, a few users get remarkable results, and most still ask simple questions.