When Claude Builds Claude: Anthropic's Recursion Report
Anthropic has published "When AI builds itself", a report on how much of its own development now runs through Claude. The headline number: as of May 2026, more than 80% of the code merged into Anthropic's codebase was written by Claude, up from the low single digits before Claude Code launched in February 2025.
The report matters less for that statistic than for what Anthropic does with it. It argues that AI systems are approaching the point where they could meaningfully help design their successors, and that the world should keep the option to slow down or pause frontier development if that happens. That is not a call to stop today. It is a lab with the top models on most leaderboards saying the brakes should be built before they are needed.
The numbers Anthropic published
- Code authorship: more than 80% of merged code authored by Claude as of May 2026.
- Throughput: in Q2 2026, the typical Anthropic engineer merged 8x as much code per day as in 2024.
- Task horizon: the length of tasks Claude completes reliably on its own has been doubling roughly every four months, faster than the earlier seven-month trend. Claude 3 Opus (March 2024) handled tasks of about four minutes; Claude 3.7 Sonnet, a year later, about 90 minutes; Claude Opus 4.6, about 12 hours.
- External measurement: METR found Claude Mythos Preview could work for "at least" 16 hours, at the upper end of what its current task suite can measure.
- Examples: in April 2026, Claude shipped more than 800 fixes that reduced a class of API errors by a factor of 1,000. On the most open-ended internal tasks, its success rate reached 76% in May 2026, up 50 percentage points in six months.
These are self-reported and hard to verify from outside. "80% of merged code" says nothing about how much human review, rework, or direction went into it. Still, the direction of every metric is the same, and the METR figure comes from an outside evaluator.
Where Claude still falls short
Anthropic is specific about the gap. Claude matches or beats skilled engineers at executing well-specified work, but "large performance gaps persist when it comes to Claude exercising judgement in choosing goals in both engineering and research." Deciding which experiment to run and which problem is worth solving is still human work. The report describes that as the gap between today's systems and one that could design its own successor.
That framing is useful for anyone deploying agents. Current models are strong executors and weak at choosing what to do. Pipelines that keep humans on goal-setting and hand execution to the model are consistent with Anthropic's own internal data.
What Anthropic actually proposes
Some coverage summarized the report as a call for a global pause. The text is narrower. Anthropic writes that if it were possible to slow the technology down to buy time, "that would likely be a good thing," and argues for preserving "the option to slow or temporarily pause frontier AI development." It says it would slow down or pause "if other developers at or near the frontier also did so in a verifiable manner."
That is a conditional commitment plus a request for verification mechanisms that don't yet exist. Whether those mechanisms ever materialize is a policy question. The factual point is that the lab publishing these numbers thinks the coordination problem is urgent.
What it means for model rankings
Anthropic's models hold the top spots on LMRank: Claude Opus 4.8 (9.7) and Claude Opus 4.7 (9.6). If model-assisted development compounds, the labs with the best coding models improve fastest, and leads could widen rather than shrink. That runs against this year's other trend, cheap open-weight models closing the gap with the frontier: see DeepSeek V4 Pro at 9.0. Both can be true for a while. Which one dominates is the question worth tracking.
Takeaway
Read the report as data from the lab that measures this most directly: agent task horizons are growing fast, execution is largely solved for well-specified work, and judgment is not. Design your own agent workflows around that split, and expect the horizon numbers to keep moving.
Related: Best Overall Models · Best Agentic Models · Best Coding Models