Anthropic: Claude Writes Over 80% of Its Own Code — and Calls for a Pause Button
On June 4, 2026, Anthropic published a report titled “When AI builds itself” and, for the first time, put concrete numbers on its own code production: in May 2026, the company says, more than 80 percent of the code merged into production came from Claude. Counting scripts and experimental code, Anthropic puts the share above 90 percent. In the same breath, the authors — Anthropic researchers Marina Favaro and Jack Clark — argue that the world should keep open the option to slow down or temporarily pause frontier AI development.
What Anthropic actually reports
- More than 80% of the code merged into production in May 2026 came from Claude; including scripts and experiments, over 90%.
- Anthropic’s engineering team ships roughly eight times as much code per day as in 2024 (a figure Anthropic itself qualifies).
- The duration of tasks a model can complete autonomously is currently doubling about every four months — down from roughly seven.
- Recursive self-improvement has not yet been reached, but “could come sooner than most institutions are prepared for.”
- Anthropic wants the option of a verifiable, multilateral development halt — not a unilateral shutdown.
What applied before
That AI coding assistants are productive has been consensus for two years. Until now, though, the internal share was an estimate from interviews — claims like “half” of the code now coming from models. There was no robust figure published by the company itself.
The debate around “recursive self-improvement” — the point at which an AI system improves itself without direct human oversight — also stayed largely theoretical. It was an argument in safety papers, not an observation from a lab’s own operations.
What applies now
1. An official number instead of an estimate. Anthropic names a concrete figure for its own operation for the first time: over 80 percent of merged production code in May 2026. This is not a vendor benchmark on a product, but self-reported data on its own codebase — to be read accordingly, but it comes first-hand.
2. Acceleration, not just level. Anthropic points to measurements by the independent organization METR: the length of tasks a model can sustain autonomously is currently doubling about every four months instead of every seven, as before. From roughly four minutes (March 2024) to about 1.5 hours (March 2025) to twelve hours today; in METR tests, a preview model reportedly sustained at least 16 hours. Read these as a trend, not a guarantee.
3. A pause mechanism as a proposal. Favaro and Clark argue that a worldwide frontier slowdown “would likely be a good thing” — but only if US and Chinese labs (and others near the frontier) stop together under rules outsiders can verify. A unilateral halt by a single lab is explicitly rejected; it would only postpone development, not slow it.
Context
What stands out about the report is the tension inside it: the same company whose model now writes most of its code is calling for an emergency brake on exactly that dynamic. You can read this as a serious safety warning — or as the positioning of a lab establishing itself early as a responsible actor in a coming regulatory debate. The two are not mutually exclusive.
When assessing the numbers, sobriety pays off. “Claude writes 80% of the code” does not mean 80% of engineering work is automated: whoever prompts, reviews, and merges the code is still human — and what’s measured is lines, not decisions. Anthropic itself qualifies the 8x productivity figure as possibly overstated. That, too, is a signal: the company frames the numbers as indicators, not proof.
The hard core remains the trend. If the doubling time for autonomous tasks really has dropped from seven to four months, the question shifts from “whether” to “when” — and that is exactly what the pause proposal targets. That recursive self-improvement is “not yet” reached is the most important caveat in the whole report: Anthropic describes an approach, not a breakthrough.
What you can do now
If you use AI coding productively: Treat the 80-percent figure as a benchmark for what’s possible, not a target. The leverage is in review, testing, and clear tasks — not in maximal auto-generation. The cascading-failures episode from April shows how quickly unnoticed regressions arise when control slips.
If you follow the safety debate: Separate the documented numbers (code share, METR doubling) from the political demand (the pause option). The former is self-reporting plus independent measurement; the latter is a position open to debate.
If you advise clients as an agency: “Recursive self-improvement” will now show up in client conversations. The sober answer: not yet reached according to the vendor, but the trend is real and measurable.
Related news
How quickly unnoticed errors arise in agentic systems is something Anthropic documented itself back in April.
→ Anthropic explains: Why Claude Code performed worse for weeks
Entdecke mehr
AI Workflows by Keyword: How We Make Recurring Routines Enforceable
A typed keyword triggers a fixed AI routine — and every single step must be committed before the next one appears. Why that's the actual trick.
GlossarEnsemble / Multi-Model Orchestration
Ensemble means combining several deliberately varied LLM runs or models whose findings complement each other. Multi-model orchestration drives these runs via orchestrators with sub-agents, so the union of results is larger than any single run.
LexikonFunction Calling / Tool Use
How an LLM uses tools: define a tool as a schema, the model picks the function and arguments, the result returns to the chat — the basis of every agent.