Anthropic data shows Claude Code success rates plateauing around 90% across task complexity
ronbodkin · x · 2026-11-20
Citing Anthropic's 'When AI builds itself' report (Sept update), ronbodkin notes Claude Code session success rates have leveled off at 90% for all task complexity levels. Discussion points: the 80-90% saturation may be insufficient to close the loop on automating R&D, though the task horizon at 80% reliability has grown from 15 to 25 minutes. The report also says Anthropic engineers now ship 8x more code per quarter than the 2021-2025 average, charting progress toward recursive self-improvement and its risks.
Related event: Toby Ord: 80% reliability emerges as bottleneck for AI task length growth(3 posts)→
More from AGI Musings
- Christiano Now Puts AI Takeover Odds at 1 in 3 Within a Year — PeterBowdenLive · 2026-11-20
- DeepMind projects at least 10x annual growth in effective compute and steeply cheaper cognitive work — dfinke · 2026-11-20
- Researcher flags 'technomoral friction': Claude's firm moral stances vs. humans' face-saving lies — mattbeane · 2026-11-20
- New 79-minute review maps deep learning for gene regulation across cellular contexts, from maxATAC to Bolero — anshulkundaje · 2026-11-20
- He Flew to Athens With One Question: What Does AI Mean Where 'Know Thyself' Began — realmeetjames · 2026-11-20
- Nick Bostrom on a16z podcast: delaying superintelligence has its own costs — a16z · 2026-11-20