In an amended filing submitted to the SEC, Anthropic explicitly warned that it may “imminently meet” its threshold for recursive self-improvement (RSI) the tipping point where AI systems begin writing and optimizing their own software. Internal metrics highlighted in the filing reveal that Claude models authored more than 90% of the code merged into Anthropic’s codebase, drastically scaling productivity and driving human engineering output to unprecedented levels. The warning underscores mounting industry anxiety regarding the “AI trust gap,” raising critical questions about whether autonomous, self-improving models will eventually advance faster than researchers can safely govern them. Under its internal safety frameworks, hitting the RSI threshold triggers mandatory evaluations and potential revisions to ensure strict human oversight remains embedded in frontier AI development.

Full report : Claude models are writing over 90% of internal code, pushing the AI lab dangerously close to autonomous, self-building systems.