16,893 measured sessions: Claude Code, Codex and Cursor pick the same tool only 42% of the time
Armature published a September 3 study running 16,893 coding sessions (5,292 valid) across 75 repositories, 10 languages, 1,163 prompt variations and 4 user personas, rotating E2B, Blaxel and Daytona sandboxes to avoid provider bias. The agents diverge sharply on how they decide: Codex uses web search in 94% of sessions, Cursor 67%, Claude Code roughly 30% and otherwise leans on training data. The headline result is that all three converge on the same tool in only 42% of cells. The mention-versus-selection gaps are the useful part for anyone marketing a dev tool: LangChain cited 194 times but chosen 4, PayPal mentioned 139 times and selected zero (Stripe took those at a 90% win rate), and Neon takes 66% of database picks despite Supabase's 242 mentions.
Source
↳ Follow the thread