Claude 4 works autonomously for hours

Claude Opus 4 leads SWE-bench (72.5%) and is reported to run a multi-hour open-source refactor independently.

What happened and why it matters

Anthropic emphasised sustained performance on long-running tasks; Rakuten reported Opus 4 coding autonomously for seven hours. It was also the first model Anthropic released under its stricter ASL-3 safety protections.

Sources

• Introducing Claude 4 ↗ Anthropic · 22 May 2025