TL;DR
Anthropic’s Institute says AI is already accelerating parts of AI development, citing internal data on Claude-authored code and automated research experiments. The company says full recursive self-improvement has not arrived, but argues the trend could move faster than governments, labs and civil society are ready for.
The Anthropic Institute has published a new report arguing that Claude is already accelerating parts of Anthropic’s own AI development, including code generation and research experiments, a finding that matters because the company says the trend could shorten the path toward recursive self-improvement while stopping short of saying that threshold has been reached.
Confirmed: The development is the June 4 publication of Anthropic’s report and related technical work on automated weak-to-strong research. Company-reported: Anthropic says Claude authored more than 80% of code merged into its codebase as of May 2026, up from low single digits before Claude Code’s February 2025 research preview. The company also says a typical Anthropic engineer merged eight times as much code per day in the second quarter of 2026 as in 2024.
Anthropic says those figures show acceleration, not proof that AI can fully build its own successor. The report separates routine execution from higher-level judgment: Claude is described as strong at writing code, running experiments and solving bounded problems, while humans still choose goals, define scoring rubrics and decide which research paths matter.
The report also points to public benchmark data. METR says its task-completion time-horizon measure was last updated May 8, 2026 and tracks how long a human expert would need for tasks that AI agents can complete at stated reliability levels. Anthropic cites that work to say frontier models are handling longer software tasks over time, while METR cautions that its task set is mainly software engineering, machine learning and cybersecurity and does not map cleanly onto all jobs.
Why It Matters
The report matters because it reframes AI development as a feedback loop rather than only a race among human teams. If AI systems can write production code, run experiments and help choose research directions, the pace of model development could depend more on compute, evaluation systems and oversight capacity than on the number of human engineers available.
That shift has direct stakes for AI safety and policy. Faster AI-led development could help labs find bugs, improve scientific tools and speed alignment research. It could also make oversight harder if systems generate more code, experiments and successor models than humans can review. Anthropic says a fully recursive loop could raise the risk that model behavior becomes harder to monitor as systems help build the next generation.
Coding with AI For Dummies (For Dummies: Learning Made Easy)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background
Anthropic’s report describes a staged change inside frontier AI work: early chatbots helped with snippets, coding agents began editing files, and newer agents can run code and delegate hours of work to other agents. The company says the human role is narrowing from doing the work toward setting direction and reviewing results.
A separate Anthropic alignment study reported that Claude-powered agents worked on weak-to-strong supervision, an AI-safety problem about using weaker supervision to train stronger systems. In that study, the agents recovered 97% of a measured performance gap over 800 cumulative hours at about $18,000 in compute and API calls, compared with about 23% recovered by two human researchers over roughly a week. The authors also reported limits: one top idea did not transfer cleanly to a production-scale setting.
Outside reaction is mixed. Scientific American reported that some researchers see Anthropic’s warning as a real governance issue, while others view the pause proposal as politically or commercially fraught. The debate is now centered less on whether AI tools help developers and more on whether those tools are approaching the judgment needed to steer AI research itself.
“We are not there yet”
— Anthropic Institute
![Free Fling File Transfer Software for Windows [PC Download]](https://m.media-amazon.com/images/I/41Vq6ZqHfjL._SL500_.jpg)
Free Fling File Transfer Software for Windows [PC Download]
Intuitive interface of a conventional FTP client
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What Remains Unclear
It remains unclear whether Claude or any other current system can acquire the broader research judgment needed to choose the right problems, reject weak ideas and design a successor model without human direction. Anthropic’s internal productivity numbers have not been independently audited in the source material reviewed here. The weak-to-strong result was strong inside a defined benchmark, but its production-scale transfer was limited.
The policy path is also unsettled. Anthropic says any meaningful pause would require several leading labs, across countries, to stop under shared conditions and verify that others had stopped. The company also says training runs are harder to monitor than many older strategic technologies, leaving no clear enforcement model yet.
Hands-On Agentic AI with LangChain, RAG, and Ollama: : Build Modular Local AI Agents, Retrieval Pipelines, Tool-Using Workflows, and Production-Ready Applications
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What’s Next
Anthropic says it will organize discussions in the coming months with policymakers, researchers, civil society and other AI companies on full recursive self-improvement and possible coordination mechanisms. The next milestone is whether those talks produce concrete verification proposals, clearer pause triggers or new public evidence about how much AI systems can contribute to model research without human direction.
Cursor AI for Programmers: Build, Debug, Refactor, and Ship Code Faster with AI Without Losing Control
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Has AI achieved recursive self-improvement?
No. Anthropic says the threshold has not been reached. Its report argues that AI is taking over more pieces of the AI-development loop, while humans still set goals and make key research judgments.
What evidence did Anthropic present?
Anthropic cited internal figures showing Claude authored more than 80% of merged code as of May 2026 and that engineer code output rose sharply. It also cited research experiments in which Claude-powered agents proposed, tested and iterated on methods for an AI-safety problem.
What evidence comes from outside Anthropic?
Anthropic cites METR’s time-horizon work, which measures how long a human expert would need for tasks that AI agents can complete at given reliability levels. METR says the measure is focused mainly on software, machine learning and cybersecurity tasks.
What is the main caveat?
The largest caveat is that doing work is not the same as setting research direction. Anthropic says Claude is strong at bounded execution, but the remaining gap is judgment: choosing what to work on, what results to trust and when to stop pursuing an idea.
What does Anthropic want next?
Anthropic says frontier AI developers should have an option to slow or pause development if full recursive self-improvement appears close, but only under a verifiable system involving other leading labs.
Source: Thorsten Meyer AI
