TL;DR
Anthropic Institute has published a report arguing that AI systems are already accelerating AI development inside Anthropic. The company says Claude authored more than 80% of merged production code as of May 2026 and that agents outperformed humans in one bounded AI-safety research task, but full recursive self-improvement has not been shown.
Anthropic Institute has published a report saying its Claude systems are now doing a large share of Anthropic’s own AI development work, including more than 80% of merged production code as of May 2026, a finding the company says points toward the possibility – not the arrival – of recursive self-improvement.
The report, When AI builds itself, was co-authored by Marina Favaro and Jack Clark and uses public benchmarks and internal Anthropic measurements. Anthropic says engineers on average now ship 8x as much code per quarter as they did from 2021 to 2025, while Claude moved from low single-digit authorship of merged code before the February 2025 research preview of Claude Code to more than 80% in May 2026.
The company also cites external evidence from METR: the length of software tasks AI systems can reliably complete on their own has been doubling roughly every four months. Anthropic says Claude Opus 3 handled roughly four-minute tasks in March 2024, Claude Sonnet 3.7 handled about 1.5-hour tasks a year later, Claude Opus 4.6 handled 12-hour tasks in March 2026, and METR found Claude Mythos Preview could work for at least 16 hours on its test suite.
On research work, Anthropic reports that Claude Mythos Preview achieved about a 52x speedup on a fixed model-training optimization task in April 2026, up from about 3x for Claude Opus 4 in May 2025. In a separate April 2026 AI-safety project on weak-to-strong supervision, Anthropic says Claude-powered agents recovered 97% of the gap between a weak-supervisor floor and strong-model ceiling over 800 cumulative agent hours and about $18,000 in compute, while two human researchers recovered about 23% over roughly a week.
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.
Coding with AI For Dummies (For Dummies: Learning Made Easy)
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.
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.
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 matters because it shifts the debate over recursive self-improvement from a purely future scenario to a measurement question inside today’s AI labs. Anthropic is saying AI has already moved from writing snippets to running parts of the engineering and experimental loop, which could shorten model-development cycles and change how labs, governments and safety teams monitor frontier systems.
The impact is not limited to Anthropic. If similar gains spread across AI companies, the constraint in AI development may move from writing code and running tests to selecting goals, reviewing outputs and verifying safety. That would put more pressure on review systems, audit tools and coordination between labs before any system can design and train its successor with little human involvement.
Background
Anthropic frames AI development as a loop with four main stages: writing code, running experiments, proposing experiments and setting direction. The company says Claude is strong at the first two and improving on the third, but human researchers still set the main problem, decide what evidence counts and decide when a path is worth pursuing.
That boundary is central to the report. The April 2026 research-agent result was open-ended inside a human-made frame: people chose the AI-safety problem and wrote the scoring rubric, while agents proposed and ran every experiment. Anthropic says this is evidence of progress toward automated AI research, not proof that Claude can independently decide what AI system should be built next.
“We are not there yet, and recursive self-improvement is not inevitable.”
— Anthropic Institute
“The comparative advantage of humans as of right now is still in seeing the bigger picture.”
— Anthropic employee, quoted by Anthropic
What Remains Unclear
Several points remain unresolved. Anthropic’s code authorship, 8x productivity figure and internal research-task results are self-reported; they are not audited in the source material. The company also says lines of code can overstate productivity, employee estimates may be high and the April 2026 agent research result did not transfer cleanly to production-scale models.
It is also unclear whether current training methods can produce the research judgment needed for a system to choose goals, evaluate its own successors and run the whole model-development cycle. Anthropic presents recursive self-improvement as plausible, not confirmed.
What’s Next
Anthropic says it will hold conversations in the coming months with policymakers, researchers, civil society and other AI companies on full recursive self-improvement and possible coordination mechanisms, including ways to verify a slowdown or pause in frontier AI development. The next milestone is whether outside researchers can validate the internal data and whether new benchmarks can measure longer, more open-ended AI research tasks.
Key Questions
Has Anthropic shown that recursive self-improvement is happening?
No. Anthropic says AI is already speeding AI development, but the report states that full recursive self-improvement has not arrived and may not happen.
What is confirmed right now?
Confirmed facts include that Anthropic Institute published the report and attributed it to Marina Favaro and Jack Clark. The internal figures are Anthropic’s own reported measurements and should be treated as company claims unless independently verified.
Why does Claude writing Anthropic code matter?
If AI systems write, test and review more of the code used to build future AI systems, model development could move faster and make human review the bottleneck.
What remains unknown?
It is unclear whether Claude or similar systems can develop durable research judgment, transfer bounded research wins to production-scale models, or make safe choices in a self-directed model-development loop.
Source: Thorsten Meyer AI
