80%+
of code merged into Anthropic’s codebase authored by Claude, as of May 2026
AGI Made Clear › AI building AI
Is AI already helping to build better AI?
Yes, and the companies doing it have started publishing figures. The same documents are careful to say something else: fully autonomous recursive self-improvement is not happening. Both halves of that are worth taking seriously.
For most of the time people have argued about AI improving AI, there was nothing to measure. The argument was about whether a feedback loop could exist in principle.
That changed during 2026. Two frontier labs published accounts of how much of their own AI development work is now done by AI, and both drew the same line in the same place.
Recursive self-improvement sounds more complicated than it is. The idea is that an AI helps create a better AI, which then helps create an even better one, potentially producing a feedback loop in which AI development itself gets faster.
We are not there. But AI is now doing enough of the work involved in developing AI that the earlier stages of that process can be measured, which is new.
What Anthropic published about its own codebase
Anthropic’s piece on recursive self-improvement, written by Marina Favaro and Jack Clark, says that “as of May 2026, more than 80% of the code we merge into Anthropic’s codebase was authored by Claude”, and that leadership estimates the figure at 90% or more once scripts and experimental code are included.[Anthropic]
The same document reports that in the second quarter of 2026 “the typical engineer was merging 8× as much code per day as they were in 2024”.
And then it discounts its own headline. Lines of code, it says, “is an imperfect measure, as it measures quantity over quality”, so the eight-fold figure “is almost certainly an overstatement of the true productivity gain”.
Published by the same company, in the same document
8×
more code merged per engineer per day in Q2 2026 than in 2024
The company that published the second number also wrote that it is “almost certainly an overstatement of the true productivity gain”. Both figures count output. Neither counts whether the right thing was built.
That caveat is the reason the document is worth reading rather than quoting. Writing most of the code doesn’t mean the AI is deciding what should be built. Humans may still be choosing the problem, setting the direction and deciding which ideas are worth pursuing. That is the part that would have to change for the loop to close.
The measurement that gets closer to the interesting question
Anthropic also reports an attempt to measure research judgement rather than research output. Given a set of candidate directions, how often does the model pick the one a human researcher would have picked?
“On this measure, our best model in November 2025 (Opus 4.5) beat the human choice 51% of the time; in April 2026 (Mythos Preview), this grew to 64%.”[Anthropic]
That matters more than the code figure, because choosing what to investigate is part of doing research rather than carrying out a researcher’s instructions.
Two things are true about that number at once. It moved 13 percentage points in five months, which is fast. And it is a comparison against one human choice on a problem somebody else had already framed, with a scoring rubric somebody else had already written. The hard part of research is upstream of both.
The same document says so: “the comparative advantage of humans as of right now is still in seeing the bigger picture and thinking beyond the confines of the immediate task”. Its verdict on the loop closing is “we are not there yet, and recursive self-improvement is not inevitable”.
The experiments behind the numbers
The 26% figure comes from an internal R&D Automation Index, which scores research tasks by how much of the work AI does, from limited assistance at the bottom to a level where a model could run without a person in the loop at the top. As of August 2026 Anthropic places Claude as leading 26% of measured work, and says no measured category has reached that final, fully autonomous level.
Some of the underlying tests are more concrete. In one recurring experiment Claude is handed the code that trains a small model and asked to make training faster without breaking it. In May 2025 Claude Opus 4 managed roughly a three times speed-up; by April 2026 an internal research model reached around 52 times. Anthropic estimates a skilled human researcher would need four to eight hours to reach about four times.
Given an open-ended AI safety research problem, two human researchers working for about a week recovered roughly 23% of the available improvement, while teams of Claude agents recovered 97%. Anthropic is careful that the result did not transfer cleanly to production scale and used substantial compute.
It has also started measuring research judgement. Shown real sessions where a human took an unproductive detour and asked what it would do next, Claude's best model chose the better direction in 51% of those selected cases in November 2025, rising to 64% by April 2026. The test is built around decisions that had room to improve, so it is not a general claim that Claude has better judgement than the researchers.
OpenAI drew the same line, three weeks later
On 9 September 2026, OpenAI’s chief global affairs officer Chris Lehane published a policy piece containing an unusually direct statement. Fully autonomous recursive self-improvement, which he defined as AI systems independently driving successive generations of increasingly capable AI, is “not happening today”.[OpenAI]
In the same piece: “AI is already accelerating parts of the research used to develop and align the next generation of models”, and “AI agents can perform some tasks that would take skilled researchers several days. This is not recursive self-improvement, but it is evidence of the direction of travel.”
On the sourcing
Both companies volunteered the capability figure and both volunteered the limit. That is better practice than the field usually manages. It is still each company measuring itself, on measures it chose, and publishing the result.
The independent read is that nobody can settle it yet
METR published an assessment in July 2026, written by Parker Whitfill and Tom Cunningham with seven other economists, which is the closest thing to a neutral position on the question.[METR]
Its conclusion is that we do not yet have enough evidence to know what would happen if AI became substantially better at AI research.
Plenty of things could slow the loop down. AI development still depends on data, on computing power for training and for running the models, on experiments that take real time, and on physical infrastructure. Better AI research does not automatically remove any of those.
But the authors also concluded that the evidence is not strong enough to say any one of those constraints would definitely prevent rapid acceleration. In their words, they “cannot rule out an extended and rapid acceleration in capabilities”, and the evidence for the constraints is “not overwhelming” either.
So the independent answer is frustratingly simple. We do not know yet. The paper is explicit about why: “there is more data relevant to RSI that the labs could be releasing.”
What is actually established, and what is not
EstablishedAI writes most of the code at at least one frontier lab, runs experiments, and is measurably better than it was at choosing between research directions somebody else has framed.
Asserted by the people measuring itThat humans still hold the advantage in framing the problem, and that the productivity figures overstate the real gain.
Not established either wayWhether the feedback effects are strong enough to compound, and which bottleneck binds first. The independent assessment says the evidence is thin in both directions.
Not happeningA system independently driving successive generations of more capable systems. Both labs that have addressed it directly say so.
Why this sits underneath the slowdown argument
The case for deliberately slowing frontier development rests on capability moving faster than the ability to evaluate and control it. AI-assisted AI research is the mechanism by which that could happen, because it is the one input to AI progress that is itself an AI capability.
That is why the distinction between the stages is not pedantry. AI making researchers faster changes how quickly models arrive. AI choosing what to research changes who is steering.
01
AI does the workHumans decide what to research. AI helps carry it out.
02
AI chooses the researchAI starts deciding which ideas are worth pursuing.
03
AI closes the loopAI helps design and develop a more capable system, which then contributes to developing the next one.
The evidence we have today is strongest for stage one. There are early signs relevant to stage two. Stage three has not been demonstrated.
What would count as the next step
Not a larger share of the code. Four things, none of which has been reported:
The agenda, not the taskA system choosing which weaknesses of the current generation are worth working on, rather than picking well among options a human listed.
Spending decisionsA system deciding how much compute a line of research deserves. Resource allocation is where research judgement becomes expensive.
A closed loopA generation designed, trained, evaluated and released with the human contribution reduced to sign-off. Every lab currently stops short of this deliberately.
Independent verificationAll of the above measured by somebody other than the company reporting it. This is the one that has not started.
Until then, the honest description of 2026 is the one both labs used. AI is accelerating parts of AI research. It is not building its successor.
Can AI improve itself? The plain-English version →