Anthropic has revealed that AI models are actively driving their own development, with Claude authoring over 80% of code merged into the company's codebase as of May 2026. Internal report data shows Anthropic engineers merged 8x as much code per day in Q2 2026 compared to 2021–2024 baseline levels. The technical acceleration provides real-world evidence for early recursive self-improvement, where AI systems assist in building their own successors.
Accelerating Code Production and Autonomous Debugging
For much of AI development history, human engineers manually specified and wrote code. Anthropic reports a major structural shift that accelerated following the launch of Claude Code in February 2025. As of May 2026, more than 80% of code merged into Anthropic's codebase was authored by Claude, compared to low single digits in early 2025.
This transition allows agents to run code and delegate work autonomously over extended timeframes. In April 2026, Claude independently shipped over 800 fixes to resolve an API error class, cutting errors by a factor of 1,000. Anthropic estimated this single effort saved roughly four human-years of tedious debugging. Furthermore, Claude's success rate on open-ended, complex tasks reached 76% in May 2026, up 50 percentage points over six months.
Benchmarks Show Rapid Scaling in Task Duration
Beyond internal development metrics, public and benchmark evaluations highlight how rapidly AI models are extending their operational duration. The length of tasks Claude can complete reliably has doubled roughly every four months, moving from short scripts to multi-hour autonomous workflows.
- 4-minute tasks: Claude Opus 3 handled four-minute software tasks in March 2024.
- 1.5-hour tasks: Claude Sonnet 3.7 handled 90-minute tasks by March 2025.
- 12-hour tasks: Claude Opus 4.6 managed 12-hour tasks by March 2026.
- 16-hour limits: METR evaluations showed Claude Mythos Preview working continuously for at least 16 hours.
Benchmark saturation has followed a similar trajectory. SWE-bench, which evaluates software engineering fixes on real open-source codebases, moved from low single-digit scores to saturation within two years. Similarly, CORE-Bench research reproduction scores jumped from 20% in 2024 to saturation fifteen months later.
Human Judgment and Productivity Limitations
Despite these rapid gains, Anthropic notes important caveats regarding output metrics and full self-improvement. Merged lines of code per engineer per day expanded 8x in Q2 2026, but the company acknowledges that volume does not perfectly equate to quality, meaning true productivity gains are likely lower.