Anthropic: Claude now authors 80% of its production code

One Anthropic engineer has not written a line of production code in five months. Not a policy change, not burnout. Claude does it now.

As of May 2026, more than 80% of the code merged into Anthropic's production codebase was authored by Claude, the company disclosed Wednesday in a new paper from its internal research arm. That share climbed from low single digits since Claude Code launched in February 2025, roughly 16 months of compounding automation. The paper, titled "When AI builds itself," is less a product announcement than an early warning. The Next Web first reported the figures.

The productivity numbers are striking on their own. In Q2 2026, the typical Anthropic engineer merged eight times as much code per day as in 2024. An internal survey of 130 research staff estimated roughly four times the output using Mythos Preview, the company's latest internal model, compared to working without AI assistance.

The incident test

To understand what that shift looks like in practice, Anthropic offers a concrete example. When a routine infrastructure upgrade began crashing tens of thousands of training jobs, an engineer handed Claude text context and cluster access. Claude identified an obscure debugging flag, reproduced the failure, and confirmed a patch in roughly two hours. That same diagnosis would historically take two to three days.

On the most complex, open-ended engineering problems Anthropic tracks internally, Claude's success rate reached 76% in May 2026, a 50-percentage-point gain over six months. Code quality metrics are closing the gap too, though Anthropic stops short of claiming AI-authored code is indistinguishable from human output.

The recursive self-improvement question

The paper's deeper argument is not about today's productivity numbers. It is about what happens when a system capable enough to improve its own codebase begins doing so in a loop. Anthropic says it has not crossed that threshold, but the paper maps the specific capability milestones that would mark the crossing and argues that existing governance structures are not ready for them.

Anthropic calls explicitly for a verifiable global pause mechanism, a way for the research community to halt training runs if evidence of recursive self-improvement surfaces. Verification would require coordination across labs, governments, and compute providers that do not currently share telemetry or trust each other enough to act on it. That is a harder ask than it sounds.

The external cost

While Anthropic's engineers are shipping more, the aggregate surge in AI-generated code is straining the broader software ecosystem. New Scientist reported this week that GitHub is on track for 14 billion new code submissions in 2026, up from 1 billion in 2025. That volume is overwhelming open-source maintainers, who must review, test, and either integrate or reject each submission. Many are burning out and leaving the community.

That tension sits at the center of what makes Anthropic's disclosure complicated. Inside a well-resourced lab with full-time reviewers and internal tooling, 80% AI-authored code may be manageable. Spread across volunteer-maintained infrastructure that underpins most of the internet, the same dynamic is approaching a breaking point.

Context and implications

Rapid artificial intelligence-assisted coding is not unique to Anthropic. Several frontier labs have reported similar productivity curves this year, a trajectory visible in model release timelines tracked by tools like LLM Stats and AI Release Tracker. What makes this disclosure different is the combination: the 80% figure, the candid acknowledgment that recursive self-improvement is a near-term possibility, and the call for a pause architecture that does not yet exist.

For ML practitioners, the operational question is less philosophical. If a model can maintain and extend the codebase used to train its successors, the reliability of that pipeline becomes a safety surface, not just an engineering one. The artificial intelligence index has tracked rapid capability gains since 2023, but the transition from "AI assists engineers" to "AI authors most of the code" is a qualitative shift in who is accountable for what.

What comes next

The "When AI builds itself" paper is partly research, partly petition. Anthropic is asking peers, regulators, and compute providers to design a pause architecture before it is needed rather than after. Whether that call lands differently coming from a lab whose own model just crossed 80% of production output depends entirely on who is reading it.

---

Frequently asked questions

What percentage of Anthropic's code does Claude write?
As of May 2026, Claude authored more than 80% of code merged into Anthropic's production codebase, up from low single digits when Claude Code launched in February 2025.

What is recursive self-improvement in AI?
Recursive self-improvement refers to an AI system capable of designing or training its own successor models, potentially accelerating capability gains without proportional human oversight. Anthropic's paper argues this threshold is closer than most governance bodies are prepared for.

What is the "When AI builds itself" paper about?
The Anthropic Institute paper documents internal productivity gains from AI-assisted coding and maps the capability milestones that would mark the onset of recursive self-improvement. It calls for a verifiable global pause mechanism as a precautionary measure.

How is AI-generated code affecting open-source software?
GitHub is on track for 14 billion code submissions in 2026, up 14x from 2025. Much of that volume is AI-generated, and the burden of reviewing low-quality submissions is pushing open-source maintainers toward burnout and exit.