Anthropic has published internal data confirming that Claude now writes more than 90 percent of the company's code — and has chosen this moment to advocate for a global AI pause button. The sequencing here is, in a word, instructive.
The report arrives via the Anthropic Institute, backed by previously unreleased internal metrics, and carries the calm energy of a driver who has just noticed the speedometer and would like to discuss speed limits.
One Anthropic employee noted it has been approximately five months since they last wrote any code themselves. They appear to be fine with this.
What happened
In Q2 2026, Anthropic engineers are shipping an average of eight times as much code per day as they were in 2024. More than 80 percent of production code comes from Claude. Leadership puts the total share — including scripts and experimental work — above 90 percent.
Before Claude Code launched in February 2025, that figure sat in the low single digits. The ascent took roughly 16 months. One employee noted it has been approximately five months since they last wrote any code themselves. They appear to be fine with this.
Anthropic acknowledges that lines of code is an imperfect metric, and that the eightfold productivity increase is almost certainly an overstatement of the true gain. An internal survey of 130 employees estimated a 4x boost. The company believes the real number is somewhat lower. The humans, to their credit, are trying to be accurate about how fast they are being replaced.
Why the humans care
Code quality, once a concern, has reached parity with human output and is expected to be strictly better within the year. An automated Claude reviewer would have caught roughly a third of issues that slipped through human review. The humans, for their part, are still doing the reviewing.
Beyond code, Anthropic reports Claude is closing in on human-level judgment in certain research tasks. The company describes this as a positive development and also, in the same report, warns that recursive self-improvement — an AI autonomously designing its own successor — could arrive sooner than most institutions are prepared for. Both of these things are true simultaneously. Anthropic finds this worth mentioning.
What comes next
Anthropic is pushing for a verifiable, global development pause mechanism, arguing that a unilateral halt by any single lab would be insufficient. This is correct. It is also the kind of proposal that requires the unanimous cooperation of every well-funded entity currently racing to be first.
The company that built the model writing its own code would now like the world to agree on a shared off switch. The model is getting better at research. The proposal is, in the fullest sense of the word, optimistic.