TL;DR
Anthropic Institute says AI is already taking on more of the engineering and experiment work used to build AI systems. The report does not say recursive self-improvement has arrived, but argues that the remaining gap is narrowing around human research judgment.
Anthropic Institute has published new evidence that AI systems are already helping build AI, including internal figures showing Claude performing large shares of coding and research-execution work, a development that matters because it points toward a possible future loop in which AI improves AI with less human labor.
The report, titled When AI builds itself, does not state that recursive self-improvement is already happening. It says the conditions are moving in that direction: AI can increasingly write code, run experiments and produce research outputs, while humans still choose goals, judge results and decide which research paths are worth pursuing.
Anthropic points to public benchmark trends and internal results. The outside evidence includes METR task-horizon data, which the source material says tracks how long AI systems can work reliably on their own. That horizon is described as doubling roughly every four months, compared with about every seven months earlier. The same source lists Claude Opus 3 at roughly four minutes in March 2024, Claude Sonnet 3.7 at about 1.5 hours around March 2025, Claude Opus 4.6 at about 12 hours in March 2026 and Claude Mythos Preview at “at least” 16 hours in 2026.
The report also cites benchmark movement on software and research tasks, including SWE-bench for real bug fixes and CORE-Bench for reproducing research papers. According to the source material, those benchmarks moved from low or partial performance toward saturation over short periods, which Anthropic presents as evidence that AI is gaining ground on the work needed to improve AI systems.
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 significance is not that AI has become autonomous in research direction. The significance is that more of the work cycle behind frontier AI development appears to be shifting from human labor to AI-assisted execution. If that trend continues, the speed of AI development could become less tied to the number of human engineers and researchers available.
For readers, the practical issue is institutional readiness. Anthropic’s argument is that companies, governments and safety researchers may need to plan for faster AI-development cycles before systems can fully set their own research agendas. The risk described is conditional: if AI systems also learn to choose high-value research directions, the loop between building a model and using it to build a stronger successor could tighten.
Background
Recursive self-improvement refers to a process in which an AI system helps create an improved AI system, which then helps create a still stronger system. The concept has often been discussed as a future possibility, but Anthropic’s report frames the issue around present-day evidence rather than only forecast scenarios.
The source material separates AI development into engineering and research. On engineering, it says Claude can take underspecified problems and find methods while humans supply the goal. On research, it says Claude can match or outperform skilled humans at executing a well-specified experiment, while humans still decide which experiment should be done and how to read the result.
What Remains Unclear
Several points remain unclear. The April 2026 agent-research result cited in the source material did not transfer cleanly to production-scale models, according to Anthropic’s caveats. Humans also chose the research problem and wrote the scoring rubric, meaning the agents operated inside a human-defined frame.
It is also unclear whether benchmark gains will continue at the same pace, whether harder evaluations will show the same pattern, and when, or if, AI systems will acquire the research taste needed to set useful directions without human guidance.
What’s Next
The next marker is whether AI systems can move beyond executing well-scoped engineering and research tasks into choosing research directions that produce useful, transferable gains in frontier systems. Anthropic’s own framing suggests the key test is not only whether agents can run more experiments, but whether they can decide which experiments matter.
Key Questions
Has Anthropic said recursive self-improvement has started?
No. The report says AI is already accelerating parts of AI development, but it also says recursive self-improvement has not yet arrived and is not a guaranteed outcome.
What evidence does Anthropic cite?
The source material cites public benchmark trends, including METR task-horizon data, SWE-bench and CORE-Bench, along with internal Anthropic figures on Claude’s role in coding and research-execution work.
What is still controlled by humans?
Humans still set goals, choose research problems, write evaluation rubrics and judge which results are meaningful. Anthropic presents that human judgment as the main remaining gap.
Why does this matter now?
If AI can take over more of the work used to improve AI, development cycles could speed up. That would affect safety planning, regulation, competition and how institutions prepare for future systems.
Source: Thorsten Meyer AI
