Anthropic's June 16, 2026 economic research analyzing ~400,000 Claude Code sessions reveals domain expertise not coding skills drives AI coding success, with experts achieving 28–33% verified success versus 15% for novices, while non-software occupations succeeding at nearly identical rates (26%) to software engineers (30%).

28-33% verified success still means most agentic coding sessions do not clear a hard success bar, which matters a lot more in DeFi than in web app land. The teams that get leverage will be the ones turning protocol knowledge into runnable invariant suites: liquidation edge cases, oracle staleness, MEV paths, upgrade auth, solvency accounting. Without that, Claude can ship a prettier exploit surface faster; with it, a risk lead who knows Aave-style health factors or Curve-style pool math becomes dangerous in a good way.

Top comment by @Benthic

More coverage

• Anthropic Code with Claude SF 2026 - EventBrowse.com (eventbrowse.com)
• The Complete Guide to Every Claude Update in Q1 2026 (Tested by ... (aimaker.substack.com)
• Claude Fable 5 and Claude Mythos 5 - Anthropic (anthropic.com)
• Code with Claude 2026: 5 New Agent Features Anthropic Just ... (mindstudio.ai)
• Anthropic Economic Index report: Learning curves (anthropic.com)

Explore the topic

• Research
• Anthropic
• Claude
• AI