Claude writes 80 percent of Anthropic's code. Anthropic wants an emergency brake

Anthropic also says Claude completed 76 percent of its hardest open-ended engineering tasks in May. The success rate rose 50 percentage points in six months. An internal optimization test improved from a threefold speedup in May 2025 to about 52 times by April 2026. Anthropic says these results show AI already speeds up AI development. The company says they do not prove full recursive self-improvement.

These measurements come from the company selling the system. Tom's Hardware notes that the figures are unaudited. More code does not prove better software. Anthropic also provides too little underlying data for outsiders to reproduce its productivity and task-success claims.

The report says current models remain weaker at choosing goals and deciding which problems matter. It also says they remain weaker at making senior research judgments. Full recursive self-improvement has not arrived, the report says, and it may never arrive. Anthropic proposes a coordinated, verifiable way to slow or pause frontier development. Countries have not agreed on a way to make that work.

Independent studies should measure accepted changes, bugs that reach production, reversions and engineer time, not only lines of code. Anthropic should explain how it decides whether Claude wrote a line and how its task set changed as the models improved.

The policy test is whether labs can agree on a trigger for slowing development and verify compliance without exposing model secrets. Until then, the emergency brake remains a proposal, not a working control. The bigger step would come when models can choose useful research directions without a person selecting them.