"When AI builds itself" is a policy essay from The Anthropic Institute, Anthropic's unit for studying AI's effects on the world from inside a frontier lab. It asks how far AI has already gone toward recursive self-improvement, which the essay defines as "an AI system capable of fully autonomously designing and developing its own successor", and what should be done if the trend continues. Its answer to the first part is that Anthropic is already delegating a growing share of its own AI development to Claude, and that this is speeding the company up. Its answer to the second is a set of three possible futures and a conditional position on slowing down. The Institute's research agenda, published a month earlier (see Focus areas for The Anthropic Institute), had promised more detail on how Anthropic's work has sped up with AI tools and ideas about the implications of recursive self-improvement; this note reads the essay as that report.
The essay was published on 4 June 2026 (the date the Institute's index page gives). Its closing credits say Marina Favaro and Jack Clark co-authored it, with editorial support from Santi Ruiz and visuals made from data collected by Brian Calvert and Jun Shern Chan. This note follows the page as read on 12 November 2026, which carries an "Update 9/18/2026" adding an extended chart. Against the Internet Archive copy of 4 June 2026, the text differs in two other places, neither marked on the page: one staff quotation was reworded (in June, "Claude has eaten the favors. It's faster"; now "[Claude is] faster"), and one name, Chloe Lubinski, was added to the people the credits thank for feedback. The archived text does not preserve the charts, so changes to them other than the update cannot be checked. The page has no numbered sections or figures: locations are given by section heading, chart title and footnote number (footnotes 1 to 8, plus a starred note on staff quotations).
Terms used in this note
- Recursive self-improvement. In the essay, an AI system that can fully autonomously design and develop its successor. The essay says "We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for."
- Time horizon. METR's measure of the length of task, in human working time, that a model completes with 50% reliability (footnote 1).
- Benchmark saturation. A benchmark on which models score near the maximum, which footnote 2 notes is often below 100% because of errors in the questions.
- Amdahl's law. The principle that speeding up one part of a process is capped by the parts that do not speed up; the essay applies it to organizations.
The argument in outline
The essay sorts AI development into engineering (writing code, running infrastructure and training) and research (choosing experiments, interpreting results, deciding what to try next). Its claim, in the section "Evidence from within Anthropic", is that in engineering Claude can now take an underspecified problem and find the method, and in research it can "match or outperform skilled humans at executing a well-specified experiment", but that "large performance gaps persist" in choosing goals. That gap, the essay says, separates today's AI from a system that could design its own successor. It then argues that even if the gap stays open, the rest of the work being automated implies compounding acceleration ("What if we're wrong?").
The evidence the essay cites
The essay mixes public measurements with internal Anthropic data that is published here for the first time. Each item below gives what the essay says, where it comes from, and what kind of evidence it is.
Public benchmarks ("Evidence from the outside world")
- Task length. The essay says the length of tasks models complete on their own has been doubling roughly every four months, up from an earlier seven, and gives Claude Opus 3 at about four minutes (March 2024), Claude Sonnet 3.7 at about an hour and a half, and Claude Opus 4.6 at 12-hour tasks. METR's time-horizon page, as fetched on 12 November 2026, gives a doubling time since 2023 of about 129 days (about four months) on its current suite and estimates of about 4 minutes for Claude 3 Opus and about 12 hours for Opus 4.6, but about 60 minutes for Claude 3.7 Sonnet (95% interval 33 to 104 minutes), not an hour and a half. The Opus 4.6 estimate has a wide interval, from about 5 to 61 hours.
- The longest measurement. The essay quotes METR that Claude Mythos Preview could work for "at least" 16 hours, at the upper end of what METR can measure. METR's page notes that "Measurements above 16 hrs are unreliable with our current task suite."
- SWE-bench and CORE-Bench. The essay says models went from low single digits to saturating SWE-bench in two years, and on CORE-Bench from roughly 20% in 2024 to saturation fifteen months later. The CORE-Bench paper (Siegel et al., 2024) reports a best agent accuracy of 21% on its hardest tasks; this note did not check the saturation claims.
These are capability measurements on fixed task suites. They do not measure AI's effect on AI development, which the essay says requires "direct evidence from within AI companies like Anthropic".
Internal Anthropic data ("Evidence from within Anthropic")
- Share of code written by Claude. More than 80% of code merged into Anthropic's codebase as of May 2026 was authored by Claude, against low single digits before Claude Code's research preview in February 2025. Footnote 3 says leadership has publicly estimated 90% or more, and that the 80% figure counts lines merged to production attributable to Claude, with gaps in the attribution pipeline.
- Code per engineer. The chart "Code contributed per person, by quarter" plots lines merged per active contributor per day, as a multiple of the pre-2025 average, with the 99th percentile of per-pull-request line counts capped. Its printed values rise from 1.2× (first quarter of 2025) to 5.8× (first quarter of 2026) and 8.0× for the second quarter of 2026, which the caption marks as a partial quarter. The essay states the comparison three ways: "8x as much code per quarter as they did from 2021-2025" (opening), "8× as much code per day as they were in 2024" (body), and against the pre-2025 average (chart). It adds that lines of code is "an imperfect measure" and that 8× is "almost certainly an overstatement of the true productivity gain".
- Staff estimate of their own uplift. In a March 2026 poll of 130 employees, the essay says, the median respondent estimated about 4× as much output with Claude Mythos Preview as with no AI. Footnote 5 points to the Claude Opus 4.7 system card for the method. That card (Section 2.3.5) describes an informal Slack poll: "130 people reacted", the distribution was wide, the geometric mean was on the order of 4×, and "The survey was opt-in based on interest rather than a random sample"; the card treats the number as "highly uncertain". The essay itself expects "the true degree of uplift in March was somewhat lower" and, in footnote 6, cites METR's finding that developers overestimate AI's effect on their productivity (Becker et al., 2025, a randomized trial in which developers estimated a 20% speedup and were measured 19% slower).
- Session success. The chart "Claude Code session success rate" plots, for internal Claude Code sessions, the share a Claude judge deems successful without corrections, by task difficulty that an LLM classifier assigns from a session summary. On "open-ended problems" the essay reports 76% in May 2026, "up 50 percentage points in six months". The September update extends the chart to 6 September 2026, where the open-ended line reaches about 91% (read from the chart), and its caption adds that automated sessions are excluded and that most open-ended weeks in 2025 had under 100 sessions. Neither caption reports validation of the judge or the classifier.
- Retrospective code review. An automated Claude review of every change would, by Anthropic's retrospective analysis, have caught "roughly a third" of the bugs behind past claude.ai incidents. No method or count is given.
- Code quality. The claim that Claude's code is "roughly at parity" with human code is a statement of staff opinion; the essay says there "isn't full consensus" and reports no measurement.
- Anecdotes. Claude shipped over 800 fixes in April 2026 that cut a class of API errors by a factor of one thousand, where the overseeing engineer estimated a human would have needed four years; and Claude found a crashing training-job flag in about two hours, against two to three days of usual work. These are single cases with estimates by the people involved.
Research tasks
- Optimizing a training script. On Anthropic's fixed task of making a small model's training code run faster, Claude Opus 4 averaged about 3× in May 2025 and Claude Mythos Preview about 52× by April 2026, against about 4× for a skilled human in four to eight hours. The system cards give 2.993× for Opus 4 (Claude 4 system card, Section 7.3.3.4) and 51.91× for Mythos Preview (Mythos Preview system card, Section 2.3.3). The Mythos card also reports that on this task the model found a new reward hack, moving the computation out of the timed call, and that trials with validation exceptions were excluded from the scores (Section 2.3.3.1). Footnote 7 says the absolute multiple "should not be read as a real-world training speedup" and that the comparison across models is the informative part.
- The automated alignment researcher. The essay cites the April 2026 study in which Claude agents recovered 97% of a performance gap on weak-to-strong supervision, against roughly 23% for two human researchers over about a week, over 800 cumulative hours and about $18,000 of compute, and notes the caveats that the result did not transfer cleanly to production-scale models and that humans chose the problem and scoring. It concludes that "Direction-setting was the only meaningful role a human played." The library's note on that study, Automated Alignment Researchers, reports that the study's released list of human-written research directions outlined methods on which three of its five highlighted results build.
- Choosing the next step in research. From internal Claude Code research sessions (January to March 2026), the authors picked 129 moments where the researcher's next move sent the session sideways, showed models only the work before that moment, and had a separate Claude that could see how the session ended judge whether the model's or the human's next step was better. The chart "Where a researcher went wrong, could Claude have done better?" gives Claude Opus 4.5 (November 2025) 51% and Claude Mythos Preview (April 2026) 64%, with ties of 10% and 9%, against a "practical ceiling" of 90%. The essay notes that because the moments were chosen where the human had room for improvement, "this isn't a like-for-like comparison". As a check on judge bias, on 127 moments where the human's move was already strong, models were judged better about 20% of the time (footnote 8).
The three futures the essay considers ("Possible futures")
- The trend stalls, while today's capabilities spread. Exponential curves may turn out to be S-curves; research judgment may not come from scaling; compute, energy or chips may bind. Even then the essay expects large change, citing Project Glasswing's report of more than ten thousand high- and critical-severity vulnerabilities found by Mythos Preview. The authors say: "We include this scenario for completeness, but we don't believe it's likely."
- Compounding efficiency gains with humans still directing. AI development is substantially automated, but humans set directions and judge results, and bottlenecks move elsewhere, as in Amdahl's law; the essay names human code review as one Anthropic already sees. It warns that the same productivity could serve surveillance and individually tailored manipulation. The authors say "we're likely heading into this scenario."
- Full recursive self-improvement. Progress is set by compute and efficiency gains, humans move mostly to oversight and verification, and the authors say they are least certain how alignment fares: models might be aligned and wise enough to solve or halt, or rare misalignment in today's models "could compound as the models build their successors". They also argue that physical, social and institutional bottlenecks would still set the pace most people feel.
These are the authors' scenarios and judgments of likelihood; the essay gives no probabilities.
What the essay proposes ("What should we do?")
The essay states that "If it were possible to effectively slow the development of this technology to give ourselves more time to deal with its immense implications, we think that would likely be a good thing", but that a slowdown that lets "the least cautious actors" catch up could leave everyone less safe. It says it would be good for the world to have the option to slow or pause frontier development, and that the Institute will research and help build the verification systems a credible pause would need. Its conditional position is: "If such systems existed, we expect that we would slow down or temporarily pause, if other developers at or near the frontier also did so in a verifiable manner."
It lists what a pause would require: several well-resourced labs in several countries agreeing to stop under the same conditions, each able to verify the others, with defined triggers, release conditions and an adjudicator. It argues that detecting a concealed training run is harder than detecting a missile silo, that arms-control verification regimes took decades, and that "A unilateral pause by one lab, by contrast, is achievable immediately, but accomplishes much less". It commits to organizing conversations with policymakers, researchers, civil society and other AI companies in the coming months, and to publishing what comes of them.
Authorship and stake
The authors are Anthropic staff; the Institute's launch announcement, as summarized in the focus-areas note, names Jack Clark as an Anthropic co-founder and its Head of Public Benefit. The internal evidence concerns Anthropic's own product used inside Anthropic, is not released, and cannot be checked from outside. The pause position is conditional on verifiable action by other frontier developers, and the essay's argument against a unilateral pause is that it would change "who the front-runner is" without creating a wider process.
Later documents
On 17 September 2026 the Institute published Measurements for understanding the pace of AI development inside frontier labs, which links this essay and reports three measures with an Anthropic snapshot. Among them, an "Anthropic R&D Automation Index" built on an Epoch AI scale, whose ratings are made by Claude, judges that as of August 2026 Claude "leads" 26% of Anthropic's AI R&D work and operates fully autonomously on none of it. That post also says Anthropic plans to embed independent third-party evaluators at the company to verify safety practices and monitor such metrics. The essay's own page was updated on 18 September 2026 with the extended session-success chart. Among the items on the Institute's index page when fetched on 12 November 2026, none is described as reporting the outcome of the conversations the essay promised; this note did not search further.
Limitations
- Evidence strength varies by claim. The public measures are external but are capability measures. The internal measures are model-judged (session success, next-step comparisons), self-reported (the uplift poll), counted with a measure the essay itself calls imperfect (lines of code), or anecdotal. None has a published method beyond the chart captions and footnotes, and none is independently verifiable.
- Borrowed numbers differ from their sources in places. The poll is described as a median in the essay and as a geometric mean from an opt-in Slack poll in the system card it cites; METR's current estimate for Claude 3.7 Sonnet is about an hour, not an hour and a half; and the code-per-engineer comparison is stated against three different baselines.
- Judges. The session-success and next-step results are graded by Claude models. The next-step check on strong human moves addresses one form of judge bias; no human validation is reported for either.
- Selection. The next-step moments were chosen where the human erred, which the essay states, and the anecdotes are chosen examples.
- Scenarios are not forecasts. The essay's likelihood statements ("we don't believe it's likely", "likely heading into") are the authors' judgments, not estimates from a model.
Research questions
- Does Claude's share of merged code track output that matters? Lines merged per engineer is the essay's main internal series. Comparing it with measures such as shipped features, incident rates or review time over the same quarters would show how much of the 8× reflects productivity.
- How accurate is the Claude judge behind the session-success chart? Having engineers label a random sample of sessions as successful or not, by difficulty tier, would give the judge's agreement with people and show whether the rise on open-ended problems survives.
- How do staff estimates of their uplift compare with measured speedups? Repeating the poll with a random sample and pairing it with a controlled comparison on matched tasks, as in METR's trial, would test the essay's expectation that the true March figure was lower.
- Does the next-step result hold with a judge that does not know the outcome? The judge saw how each session ended. Having researchers blind to the outcome compare the two suggestions on the same 129 moments would test whether the judge's hindsight drives the result.
- What would a verification regime for a coordinated pause need to observe? The essay lists triggers, release conditions and detection of concealed training. Proposals specifying the observable signals (compute use, data center power, chip tracking) and their detectability would show how far verification is from the essay's requirements.
- Do the Institute's later measures move as the essay's scenarios predict? Tracking the R&D Automation Index and session-success series over time against the essay's second scenario, in which humans keep direction-setting, would show whether "AI leads" work grows while human direction-setting holds.
How it connects
- The Institute's agenda. Focus areas for The Anthropic Institute (May 2026) is linked from the essay and asked how to measure the speed of AI R&D and where intervention points would lie if acceleration needed slowing. This note reads the essay as the agenda's promised report on AI-driven R&D; the essay does not describe itself that way.
- The research the essay cites. Automated Alignment Researchers (April 2026) is the essay's example of an end-to-end research project. The Claude 4, Claude Opus 4.7 and Claude Mythos Preview system cards are linked for the training-speedup task and the staff poll.
- The research that cites the essay. Automated researchers can reliably mitigate alignment failures (August 2026) links the essay in its first sentence as the reason automating alignment research matters.
- Policy context. The essay links 2028: Two scenarios for global AI leadership for its point that a slowdown could let less cautious actors catch up. This note adds How AI is transforming work at Anthropic (2025), an earlier internal study of the same workplace that the essay does not cite.
References
- Favaro, M., and Clark, J. (2026). When AI builds itself: Our progress toward recursive self-improvement, and its implications. The Anthropic Institute, Anthropic, 4 June 2026 (updated 18 September 2026; read 12 November 2026). https://www.anthropic.com/institute/recursive-self-improvement record
- Internet Archive copy of the essay, 4 June 2026: http://web.archive.org/web/20260604162030/https://www.anthropic.com/institute/recursive-self-improvement
- METR. Time horizons of frontier AI models (page fetched 12 November 2026; changelog entry of 8 May 2026 on measurements above 16 hours). https://metr.org/time-horizons/ record
- Siegel, Z. S., et al. (2024). CORE-Bench: Fostering the Credibility of Published Research Through a Computational Reproducibility Agent Benchmark. arXiv:2409.11363. record
- Becker, J., et al. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR, arXiv:2507.09089. record
- Anthropic (2026). Claude Opus 4.7 System Card, Section 2.3.5 "Internal survey results (for Claude Mythos Preview)". record
- Anthropic (2025). System Card: Claude Opus 4 & Claude Sonnet 4, Section 7.3.3.4 "LLM training". record
- Anthropic (2026). Claude Mythos Preview System Card, Sections 2.3.3 and 2.3.3.1. record
- Anthropic (2026). Measurements for understanding the pace of AI development inside frontier labs. The Anthropic Institute, 17 September 2026. https://www.anthropic.com/institute/measuring-pace-of-ai-development record