TL;DR
Anthropic’s Institute says AI is already accelerating parts of AI development, citing public benchmark gains and internal data on coding and research agents. The report stops short of saying recursive self-improvement has arrived, saying humans still set goals, judge results and define research taste.
The Anthropic Institute says AI systems are already measurably accelerating AI development, citing public benchmarks and internal Anthropic data showing Claude performing more coding, experimentation and research execution work in the frontier-model pipeline. The report matters because it frames recursive self-improvement as a possible next step, while also saying that the core bottleneck — human judgment about which research directions matter — has not yet been automated.
Anthropic’s account separates what AI can already do from what remains outside its reliable reach. The confirmed claim from the source material is that Anthropic has published or described measurements showing AI taking on more of the work around AI development, including coding, running experiments and producing research results. The stronger interpretation — that this could become recursive self-improvement — is presented as a risk scenario, not as a completed development.
The report cites public METR task-horizon data showing the length of tasks AI models can complete on their own doubling roughly every four months, faster than a prior seven-month pace. The source material lists Claude Opus 3 at about 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 around March 2026, and Claude Mythos Preview in 2026 at “at least” 16 hours. It also says SWE-bench moved from low single-digit performance to saturation in two years, while CORE-Bench rose from about 20% in 2024 to saturation roughly 15 months later.
Anthropic’s internal figures, as described in the source material, include claims that Claude generated more than 80% of merged code and that code output per engineer rose eightfold. The same account describes an April 2026 test in which Claude agents ran an open-ended AI-safety research project on weak-to-strong supervision, designing experiments, testing hypotheses and sharing results across parallel agents. Anthropic reported that the agents recovered 97% of the measured gap between a weak-supervisor floor and a strong-answer ceiling, compared with about 23% for humans in a week, using about 800 cumulative agent hours and roughly $18,000 in compute.
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 is relevant because it moves the recursive self-improvement debate from theory toward measured workflow changes inside AI labs. If AI systems can cheaply perform more of the engineering and experimentation needed to build better AI, the pace of model development could depend less on human labor and more on compute, infrastructure and oversight capacity.
For readers, the near-term issue is not whether an AI system has fully taken over its own improvement. The more immediate question is how quickly labs, regulators and companies can evaluate systems that are already helping build the next generation of systems. Faster research cycles could bring useful advances, but they may also shorten the time available for safety testing, governance decisions and external review.
Background
Recursive self-improvement refers to a loop in which an AI system helps design or improve a successor system, which then becomes better at the same task. Anthropic’s report does not say that loop is operating today. It says the supporting pieces are becoming easier to observe: AI writes code, runs experiments and produces research outputs with less direct human work.
The source material describes a remaining gap between execution and direction-setting. Claude can take a defined engineering problem and find a method, and it can execute a well-specified experiment. Humans still choose the problem, decide what results mean, write evaluation rubrics and decide when a research direction is worth pursuing.
What Remains Unclear
Several points remain unclear. The internal Anthropic figures are described in the source material but are not independently verified here. The April 2026 agent result was bounded by a human-chosen problem and human-written scoring rubric, and Anthropic said the result did not carry cleanly to production-scale models. It is also unclear whether AI systems can reliably develop the research taste needed to choose valuable experiments, reject misleading results and set long-term technical direction.
What’s Next
The next milestones are harder public benchmarks, more transparent reporting from AI labs and evidence on whether agents can choose research questions, define evaluation methods and produce results that transfer to large production systems. Until then, the confirmed development is narrower: AI is doing more of the work of AI development, while humans still hold the main decision points.
Key Questions
Has Anthropic shown that AI can improve itself without humans?
No. The report says AI is taking on more coding and research execution work, but humans still select the problem, set goals, write rubrics and judge results.
What was the April 2026 Anthropic experiment?
Anthropic described Claude agents running an open-ended AI-safety research project on weak-to-strong supervision. The agents proposed hypotheses, ran tests and shared findings, but the research frame was still set by humans.
Why does the METR task-horizon data matter?
It tracks how long an AI system can work independently on a task. A faster-growing task horizon suggests AI systems may soon handle larger chunks of engineering and research work without constant human steering.
What is still missing for recursive self-improvement?
The main missing piece is reliable research judgment: choosing the right problems, recognizing strong evidence and deciding when a path is not worth pursuing.
Source: Thorsten Meyer AI
