What Happened
Anthropic published a report on Thursday that puts a number on something the AI industry has mostly discussed in the abstract. The company said its own model, Claude, now leads 26% of Anthropic's model research and development work — meaning the model takes a high-level prompt and carries most of the task through to the end, with a human still watching. At the start of 2026, that figure was effectively zero.
The report separates "leads" from a broader category. Roughly 90% of Anthropic's R&D now happens in what the company calls collaboration with Claude — the model doing large chunks of work under close human direction. The 26% figure is the narrower, more striking claim: work where the model sets the pace and a person reviews the result rather than the other way around. Anthropic was explicit that this is not autonomy. Nobody has taken the human out of the loop.
The announcement arrived in the middle of a busy stretch for this particular idea. Earlier in September, OpenAI said it had reached a target Sam Altman set publicly a year ago: an "automated research intern," which the company defines as a system able to carry out well-defined research tasks under human direction, including work that would occupy a skilled researcher for a few days. OpenAI put a second date on the board at the same time — an automated AI researcher, not an intern, by March 2028. The company also offered a metric: measured in eight-hour workdays, its research organization now consumes roughly 3.1 agent-workdays for every human workday, a ratio it says it crossed only after June.
Put those two disclosures side by side and you get the outline of what researchers call recursive self-improvement, or RSI: AI systems that help design their own successors, which in turn help design theirs. Anthropic encouraged competing labs to publish comparable numbers. As of now, it and OpenAI are the only two that have put figures on the record.
Why It Matters
The reason a percentage matters more than a press release is that RSI is a compounding process, and compounding processes are defined by their rate. If a model contributes a fixed share of the work that produces the next model, and that share grows with each generation, the interval between generations shortens on its own. Nothing about that requires a dramatic breakthrough. It only requires the trend line to hold.
There is also a mundane engineering version of the story that deserves equal weight. Most of what AI research actually consists of is not conceptual leaps. It is writing training code, wrangling data pipelines, running ablation studies, chasing down why a run diverged at step 40,000, and reading the results of experiments someone else designed. These are precisely the tasks current models handle well — bounded, verifiable, tedious, and enormous in volume. A lab that can run three times as many experiments in a week is not a lab that has discovered a new principle. It is a lab that has bought more shots on goal.
That distinction shapes what the numbers can and cannot tell us. A 26% figure describes how much research throughput the model absorbs, not how much insight it generates. Anthropic's own framing — task completion end-to-end from a prompt — is a throughput claim. Whether throughput converts into qualitatively better models is the open question, and it is the question that separates the people who find these disclosures encouraging from the people who find them alarming.
The economics point the same direction regardless. Frontier training runs are now capital projects measured against the depreciation schedules of the accelerators they run on. If model-assisted research shortens the time between a hypothesis and a validated result, the value of every hour of compute goes up. That is a strong incentive to push the share of model-led work higher, and it exists independently of anyone's view about superintelligence.
Reaction
The disclosures landed in a field that has been arguing about this scenario for two decades, and the arguing resumed immediately.
Anthony Aguirre, a physics professor at the University of California, Santa Cruz who also runs the Future of Life Institute, read Anthropic's charts as a trajectory rather than a snapshot. His concern is speed: as models take over more of the work, the loop runs at machine tempo rather than human tempo, and each turn of it arrives faster than the last. He did not hedge his conclusion, calling full autonomy of this kind "the worst idea in the history of humanity" — and noting that the labs are doing it anyway.
John Thickstun, an assistant professor of computer science at Cornell who works on controlling model behavior, offered a deflationary counterpoint. On his reading, a version of recursive self-improvement has quietly been the norm for years: researchers have long used the previous generation of models to help write the code that trains the next one. Andrej Karpathy and others have run experiments along these lines for a while. The results, Thickstun said, have been incremental rather than transformative — real gains, no creative leaps.
Both readings can be correct at once, which is part of why the debate does not resolve. Thickstun is describing what the loop has produced so far. Aguirre is describing what happens if the share of model-led work keeps climbing at the rate Anthropic just documented. The disagreement is not really about the present; it is about whether the last twelve months are a plateau or an inflection.
The timing sharpened things further. The week before, several prominent figures in the industry had signed on to a call for slowing the pace of AI development. A quantified RSI disclosure arriving days later was always going to be read as a reply, whether or not it was meant as one.
What's Next
The labs have staked out visibly different positions, and the next year will make those positions concrete.
Elon Musk has been the most forward-leaning. He said in March that for xAI's Grok models humans are gradually getting less involved in the improvement process, and that each successive model is built by the one before it — while acknowledging the process is not yet fully automated. He suggested the fully automated version might arrive by the end of this year, and said flatly that it would not be later than 2027.
OpenAI has published dates but attached conditions to them. Alongside the March 2028 target, the company said plainly that it does not yet know how to safely reach aligned, full RSI, and that it cannot assume alignment research will keep pace with capability. Its stated reason for continuing is that an automated researcher could equally be an automated safety researcher — the same tool pointed at the problem it creates. Anthropic, for its part, has said it would slow or pause development if competitors did so in a verifiable way, a condition that currently has no mechanism behind it.
Microsoft has taken a different route entirely. Mustafa Suleyman, who runs the company's AI division, has argued for what he calls humanist superintelligence — capability in service of people rather than an unbounded entity with high autonomy, carefully calibrated and kept within limits. It is a positioning statement as much as a technical one, but it marks out a distinct answer to the same question.
The practical thing to watch is whether other labs publish their own version of Anthropic's number. A single company's self-reported metric, using its own definition of "leads," is not an industry benchmark. Several companies reporting comparable figures on a regular schedule would be, and it would turn an argument about scenarios into an argument about data.
Closing Thoughts
What strikes me about this week is how ordinary the mechanism turns out to be. There is no moment of awakening in the Anthropic report. There is a percentage that went from zero to twenty-six over nine months, attached to a definition careful enough to be auditable. The most consequential technical development of the decade, if that is what this is, is being documented the way a company documents a productivity gain.
That ordinariness cuts in two directions. It is reassuring, because a process you can measure is a process you can govern, and a number published quarterly is far more useful to a regulator than any amount of speculation. It is also unnerving, because gradual change is the kind humans are worst at noticing. Nobody is going to wake up to an announcement that the threshold has been crossed. The figure will simply be 40%, then 60%, and each report will sound as measured as this one did.
The honest position is that Aguirre and Thickstun are both describing real things, and that the difference between them will be settled by observation rather than argument. If Claude-led research yields the same incremental gains that model-assisted coding has yielded for years, the alarm was premature. If the share keeps climbing and the interval between capable models keeps shrinking, the deflationary reading will look like a failure of imagination. We should know within a few model generations — and, given the trend being described here, those generations are not going to take as long as the last ones did.
한글 요약
앤스로픽이 지난 목요일 보고서를 내고, 자사 모델 클로드가 회사 모델 연구개발 업무의 26%를 주도하고 있다고 밝혔습니다. 높은 수준의 지시 한 줄을 받아 과제 대부분을 끝까지 수행하되, 사람이 감독하는 형태입니다. 올해 초만 해도 이 수치는 사실상 0이었습니다. 연구개발 전체의 약 90%는 사람의 밀착 지시 아래 클로드와 협업하는 방식으로 이뤄집니다. 앞서 9월 초 오픈AI도 "자동화된 연구 인턴"을 달성했다고 발표하면서, 2028년 3월까지 인턴이 아닌 자동화된 연구자를 만들겠다는 목표를 함께 제시했습니다. 두 발표를 나란히 놓으면 AI가 자기 후속 모델 개발을 돕는 순환 구조, 이른바 재귀적 자기개선(RSI)의 윤곽이 드러납니다.
반응은 갈렸습니다. UC 산타크루즈 물리학 교수이자 퓨처 오브 라이프 인스티튜트를 이끄는 앤서니 아기레는 이 수치를 한 시점의 스냅샷이 아니라 추세선으로 읽었습니다. 모델이 더 많은 일을 맡을수록 순환 주기가 사람의 속도가 아닌 기계의 속도로 돌아간다는 것이 그의 우려입니다. 반면 모델 행동 제어를 연구하는 코넬대 컴퓨터과학과 존 틱스턴 조교수는 훨씬 담담한 해석을 내놨습니다. 연구자들이 이전 세대 모델로 다음 세대 학습 코드를 짜는 일은 이미 수년째 관행이었고, 그 결과는 점진적 개선이었지 도약이 아니었다는 것입니다. 두 해석은 사실 서로 모순되지 않습니다. 한쪽은 지금까지의 결과를, 다른 쪽은 비율이 계속 오를 경우를 말하고 있기 때문입니다.
각 연구소의 입장도 뚜렷하게 갈라져 있습니다. 일론 머스크는 xAI의 그록에 대해 사람이 점점 덜 개입하고 있다며 완전 자동화 시점을 올해 말, 늦어도 2027년으로 제시했습니다. 오픈AI는 2028년 3월이라는 날짜를 내놓으면서도 정렬된 완전 RSI에 안전하게 도달하는 방법을 아직 모른다고 명시했고, 앤스로픽은 경쟁사들이 검증 가능한 방식으로 동참한다면 개발 속도를 늦추거나 중단하겠다는 조건부 입장을 유지하고 있습니다. 마이크로소프트 AI 책임자 무스타파 술레이만은 자율성이 무한한 존재가 아니라 사람을 위해 세심하게 조율된 능력을 뜻하는 "휴머니스트 초지능"을 내세웁니다. 앞으로 지켜볼 지점은 다른 연구소들도 앤스로픽처럼 비교 가능한 수치를 정기적으로 공개할지 여부입니다. 그때 비로소 시나리오를 둘러싼 논쟁이 데이터를 둘러싼 논쟁으로 바뀝니다.
참고 — Fortune / AP, Bloomberg, Quartz, Spectrum News, TechBriefly