TL;DR
Anthropic Institute says AI systems are already accelerating AI development, citing internal data that Claude authored more than 80% of code merged into Anthropic’s production codebase as of May 2026. The company says full recursive self-improvement has not happened and is not guaranteed, but argues that institutions need plans for a verifiable slowdown if the loop closes.
Anthropic Institute said June 4 that AI is already accelerating AI development inside Anthropic, citing internal figures that Claude authored more than 80% of code merged into the company’s production codebase as of May 2026 and helped engineers ship far more code than in prior years. The claim matters because Anthropic says the same trend, if it reaches research direction-setting, could lead to recursive self-improvement: AI systems designing and building their own successors with limited human input.
In its report, When AI builds itself, Anthropic said the typical engineer in the second quarter of 2026 was merging about eight times as much code per day as in 2024. The company attributed the rise to engineers directing and reviewing Claude-written code rather than writing most of it themselves. Anthropic also cautioned that lines of code are an imperfect productivity measure and likely overstate the true gain.
The report said Claude’s success rate on Anthropic’s most open-ended coding tasks reached 76% in May 2026, up 50 percentage points over six months, based on the company’s internal session scoring. In one cited incident, Anthropic said Claude identified an obscure debugging flag behind widespread training-job crashes in about two hours, work the company described as normally taking two to three days.
Anthropic also pointed to research evidence. In a recurring model-training optimization test, the company said Claude Opus 4 averaged about a threefold speedup in May 2025, while Claude Mythos Preview reached about 52-fold by April 2026. In a separate April 2026 AI-safety project on weak-to-strong supervision, Anthropic said agents recovered 97% of the measured floor-to-ceiling gap over 800 cumulative agent-hours and about $18,000 in compute, while two human researchers recovered about 23% over roughly a week. The company said humans still chose the problem and scoring rubric, and that the result did not cleanly transfer to production-scale models.
When AI builds itself
Anthropic is delegating a growing share of AI development to AI. Taken far enough, that points to a system that designs its own successor — recursive self-improvement. Not here yet, not inevitable. But the case isn’t speculation: it’s data on what AI is doing to AI development right now.
The curve that hasn’t bent
METR tracks the length of tasks AI can reliably complete on its own. That horizon is doubling roughly every four months — up from every seven. Anyone can check this in public data.
Task horizon — how long a job AI can handle solo
Each model handles dramatically longer tasks than the one a year before. The line keeps going up.
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.
Amazon Product 1394249136
As an affiliate, we earn on qualifying purchases.
Two kinds of work, one persistent gap
Building a frontier model splits into engineering and research. Across both, the pattern is the same — and so is the one thing AI still can’t do well.
Code, infrastructure, training
Claude can take an underspecified problem and find a method. Humans supply the goal; they no longer need to supply the method.
Which experiments, what they mean
Claude can match or outperform skilled humans at executing a well-specified experiment. But choosing which experiment still needs a human.
The same ladder Anthropic employees climb with experience
AI development tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Amazon Product B0GWR8MPR5
As an affiliate, we earn on qualifying purchases.
Watch the human share shrink, rung by rung
Walk up the four stages of AI development. At each, the human/AI split shifts — and the real internal numbers show exactly where AI has reached parity, gone superhuman, or still trails. Tap a rung.
The human role across the development loop
The doing now costs almost nothing in human time. What’s left is the deciding.
machine learning experiment hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Agents ran an open research project end to end
April 2026: the first demonstration of Claude running an open-ended research project from hypotheses to findings — on a real AI-safety problem.
Can a weaker model reliably supervise a stronger one?
Agents were left to solve it: proposing hypotheses, testing them, sharing findings across parallel agents, iterating. Measured against the gap between a “floor” (weak supervisor alone) and “ceiling” (strong model trained on correct answers).
(humans: ~23% in a week)
· ~$18,000 compute
the agents themselves
CLAUDE AI UNLEASHED From First Prompts to Pro: The Complete Guide to Claude AI for Writing, Research, Coding, and Business (The Claude AI Mastery Series)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Picking a better next step than the human
Real research sessions where a human took a wrong turn. Models saw only the work before the detour and proposed a next step; a judge that knew the outcome scored them. The day-to-day of research is this chain of next-step calls.
“Can the model pick a better next step than the human?”
Share of moments where the model’s next step was judged better. The amber line is the practical ceiling (an ideal answer that could see the whole session).
It depends on whether the trend continues — and what we do
The piece refuses a single prediction. It lays out three scenarios, and is clear about which it finds most likely.
The exponentials turn out to be S-curves
Maybe taste can’t be scaled into existence; maybe the constraint is the supply chain — chips, grid, interconnect — not intelligence. Even so, the world still changes: Glasswing’s Mythos found 10,000+ critical vulnerabilities in weeks, and a 100-person firm does the work of 1,000.
included for completeness · they doubt itDevelopment automates; humans still steer
100-person companies doing the work of tens of thousands — revolutionary, but turnable to harm (population-scale surveillance, tailored manipulation). Bound by Amdahl’s law: speeding one part shifts the bottleneck — which is exactly why human code review became Anthropic’s new chokepoint.
★ they think we’re likely heading hereAI designs and refines its own successors
Progress paced only by compute. Humans move to oversight of an expanding “virtual lab.” The future they understand least — especially whether alignment holds, or whether rare misalignments compound as models build successors, until control slips.
the one they’re most uncertain aboutBuild the option to slow down — verifiably
The piece ends on policy, not product. A unilateral pause just changes who leads; what’s missing is the ability to verify others have actually slowed.
Why a credible pause is hard — and worth building toward
A slowdown that only lets the least cautious catch up leaves everyone less safe. So the goal is the option: systems that let frontier labs verify others have genuinely stopped. Anthropic says if such systems existed and peers paused verifiably, it expects it would too.
Detection beats verification — and even that’s tough
Training runs are easier to conceal than missile silos, inputs are general-purpose, and whoever continues while others pause inherits the lead.
We’ve done it before — slowly
Regimes like the INF Treaty built verification and trust over decades. The authors’ blunt line: “We don’t have that long.”
Reading it in proportion
- This is one lab’s account of its own internal data — much previously unreported, not independently audited.
- The soft spots are stated in the original: lines-of-code overstates productivity; the self-reported 4× is probably high; the headline research result didn’t transfer to production scale; the next-step test used cherry-picked moments.
- “More autonomous” is not “fully autonomous” — every standout result still had a human framing the problem and defining success.
- That the authors surface these caveats themselves — against their own incentive — is part of what makes the document serious.
Why It Matters
The report shifts the recursive self-improvement debate from future possibility to a narrower factual issue: how much of frontier AI development is already being automated, and which parts remain human-held. Anthropic’s evidence supports a more limited conclusion than some headlines suggest: Claude is doing large amounts of engineering and bounded experimentation, while humans still set goals, decide which results to trust and judge what work is worth doing.
For readers outside AI labs, the stakes are governance, labor and safety. If AI systems keep taking over more of the AI-development loop, model progress could accelerate faster than policy, security review and organizational oversight can adapt. If the trend stalls, today’s systems still may reshape software work by letting small teams produce at much larger scale.
Background
Anthropic framed the evidence around a ladder from execution to judgment. Public task-horizon measurements cited by the company show models handling longer autonomous tasks: Claude Opus 3 managed tasks of about four minutes in March 2024, Claude Sonnet 3.7 about 1.5 hours around March 2025, and Claude Opus 4.6 about 12 hours in March 2026. Anthropic said METR found Claude Mythos Preview could work for at least 16 hours, near the upper edge of what METR could measure without harder tasks.
The company also cited public benchmarks that have moved quickly: SWE-bench, for real bug fixes, and CORE-Bench, for reproducing research. Anthropic said both went from low or modest performance to near-saturation over short periods. Those benchmark claims are public-facing context; the internal figures about code authorship, productivity, code review and research sessions are self-reported by Anthropic and have not been independently audited in the source material.
“We are not there yet, and recursive self-improvement is not inevitable.”
— Anthropic Institute
“slow or temporarily pause frontier AI development”
— Anthropic Institute
“seeing the bigger picture”
— Anthropic employee cited by Anthropic
“Zero chance there will be a slowdown.”
— Noah Giansiracusa, quoted by Scientific American
What Remains Unclear
It is not yet clear whether current training methods can produce the research taste needed for a model to choose problems, set agendas and validate successor systems without human leadership. Anthropic’s internal numbers are self-reported; some use Claude-based judging, and the source material does not show an independent audit. The open-ended safety project was bounded by a human-selected problem and rubric, and Anthropic said the result did not transfer cleanly to production-scale models.
There is also dispute over the policy response. Scientific American reported that some outside researchers view the pause proposal as unworkable or as part of industry messaging rather than an operational plan. Anthropic said any meaningful slowdown would need multiple frontier labs in multiple countries and a way to verify compliance.
What’s Next
Anthropic said it will hold conversations in the coming months with policymakers, researchers, civil society groups and other AI companies about full recursive self-improvement and coordination mechanisms. The next milestone is whether the company releases more detailed methods, whether outside researchers can validate the internal findings, and whether rival labs engage with any proposed slowdown framework.
Key Questions
What did Anthropic publish?
Anthropic Institute published a June 4, 2026 report arguing that AI is already speeding AI development inside Anthropic, using public benchmarks and internal data about Claude’s role in coding, experimentation and research workflows.
Is Claude already building its own successor?
No. Anthropic says full recursive self-improvement has not happened. Its evidence shows Claude doing large parts of engineering and bounded research execution, while humans still set goals and decide what work matters.
What is the strongest evidence Anthropic cited?
Anthropic cited several internal figures: Claude authored more than 80% of merged production code as of May 2026; engineers merged about eight times more code per day than in 2024; Claude reached a 76% success rate on the most open-ended coding tasks; and AI agents outperformed two human researchers on one bounded safety experiment.
What remains in human hands?
According to Anthropic, humans still choose the problems, judge which results are reliable, set research direction and decide when an approach has failed. The company identifies that judgment layer as the main remaining gap between today’s systems and full self-improvement.
Why are some researchers skeptical?
The evidence is partly self-reported, some tests rely on internal scoring, and the policy proposal would require rival labs and governments to agree on verifiable limits. Critics cited by Scientific American also argue that the data may show strong productivity tools rather than a near-term self-improving AI system.
Source: Thorsten Meyer AI
