Claude Is Training Itself: Anthropic Says It Wrote Over 80% of Merged Production Code

More than 80% of newly merged lines in Anthropic’s production codebase are now written by Claude. It has also diagnosed, in two hours, a failure that crashed tens of thousands of training jobs, and designed and run experiments after being given a research question. AI is taking on the engineering and research work needed to build the next generation of AI. That raises a larger question: Could the next Claude be built primarily by Claude itself?

In “When AI builds itself,” Anthropic calls this path “recursive self-improvement.” The article traces five generations of AI development and uses public benchmarks and previously undisclosed internal data to show how Claude is speeding up coding, fault diagnosis, and experimentation. It also asks when Claude might take over choosing research questions, judging results, and designing successor models. From there, it sketches three possible futures: capability growth slows while AI spreads widely; humans remain in charge while AI keeps accelerating research; or AI completes a loop in which one generation improves the next. The article closes by asking whether laboratories could jointly slow down, in a verifiable way, if development moves too fast.

The generations of AI development

Anthropic charts five stages in its own way of working. The central change is in the size of the job people hand to AI: from a short piece of code to a goal it can break down and pursue itself.

1. 2021–2023: Human development. Engineers wrote code and documentation on their computers and drove each step of model development.
2. 2023–2025: Chatbot assistance. People asked models for code snippets and pasted them into an editor; the model could not take over a whole task.
3. 2025–2026: Coding agents. Agents began reading, writing, and changing files themselves, sometimes completing changes to an entire file.
4. Now: Autonomous agents. They can run code, test results, and delegate hours of work to other agents; people increasingly set goals and review the output.
5. Future: Recursive self-improvement. A full loop would exist if one generation of AI could design, develop, and train a successor that could then improve the next generation. Anthropic gives no date for reaching this stage.

Building a model involves two kinds of work. Engineering means writing code, building training systems, and diagnosing failures. Research means designing experiments, interpreting results, and choosing the next direction. Tasks also come in different layers: carrying out specified steps, finding a method after receiving a goal, or deciding which question is worth pursuing in the first place. Claude has moved from executing steps toward finding its own methods. Whether it can also take over judgments about direction is what matters for building its own successor.

The evidence: How AI is speeding up AI development

External benchmarks: From minutes to a dozen or more hours

The evaluation organization METR uses a “task time horizon” to track how long a job a model can complete independently. It asks how long a person would usually need for tasks on which the model reaches roughly a 50% success rate. The trend Anthropic cites shifted from a doubling about every seven months to about every four months recently. Opus 3 in March 2024 corresponded to roughly four-minute software tasks; Sonnet 3.7 a year later reached about an hour and a half; Opus 4.6 another year later reached about 12 hours. METR measured Mythos Preview at at least 16 hours, close to the limit of its testing tools at the time. Anthropic extrapolates that, if the trend holds, models might tackle tasks that take humans days later this year and tasks lasting weeks the following year.

Other tests are becoming saturated quickly. SWE-bench asks models to fix software using real repositories and bug reports; scores moved from low single digits to near saturation in about two years. CORE-Bench asks models to reproduce published results with a paper’s code and data; its success rate went from about 20% to near saturation in roughly 15 months. These tests show that AI can complete longer, more complex tasks with defined objectives. To see whether it is accelerating AI research itself, we also need to look inside the laboratory.

Internal engineering: More code, and independent fault diagnosis

As of May 2026, more than 80% of the merged lines in Anthropic’s production codebase could be attributed to Claude. Before Claude Code entered research preview in early 2025, the share was in the low single digits. The number of lines merged per engineer per day also began rising in 2025: it reached 5.8 times the earlier average in the first quarter of 2026 and 8.0 times that average in the second quarter, up to the point when the chart was made. Engineers moved from writing every line themselves toward specifying goals, reviewing output, and merging results.

That helps explain the surge in code output, while also showing that people remain in the development chain. Lines of code do not equal net productivity. As another measure, 130 research-team employees surveyed in March 2026 estimated a median productivity increase of about fourfold from their use of Mythos Preview. Automation has also made work possible that would otherwise never reach the top of the queue: in April 2026, Claude delivered more than 800 fixes that reduced one class of API errors by roughly a thousandfold. The responsible engineer estimated that doing the work manually might have taken four years.

Beyond the volume of code, the more important question is whether Claude can take on problems without a prescribed sequence of steps. Anthropic tracks internal Claude Code sessions by difficulty. By May 2026, the share of the most open-ended tasks judged by a Claude evaluator to have been completed without human correction reached about 76%, roughly 50 percentage points higher than six months earlier. Employees were also intervening to correct, redirect, or take over sessions less often. An updated chart in September put the endpoint for open-ended tasks at about 91%. That is a later observation of the same kind of internal session metric.

One concrete example is a training-cluster failure. An upgrade suddenly crashed tens of thousands of training jobs. An engineer gave Claude only a description of the problem and access to the cluster. Claude checked running jobs and environment settings one by one, located an obscure debugging flag, reproduced the failure, and confirmed the fix. It took about two hours; Anthropic estimates an engineer would normally need two or three days. The person supplied the goal, “get training running again”; Claude found the diagnostic path itself.

Once code runs, colleagues still need to understand and maintain it. Many Anthropic employees think Claude’s code has roughly caught up with human-written code, though the team does not agree unanimously. The company now also has Claude review code awaiting merger. Looking back at past incidents, it estimates that if all those changes had received automated review at the time, about a third of the bugs that later caused claude.ai incidents might have been caught beforehand.

Internal research: Given a question, Claude can experiment toward an answer

Fault diagnosis shows Claude finding its own route when the goal is clear. Research goes further: it requires proposing approaches, running experiments, and adjusting to the results. At every model release, Anthropic runs the same experiment: it gives Claude code for training a small model and asks it to make the code as fast as possible while passing correctness checks. Claude changes the code, runs it, times it, and tries again. Opus 4 in May 2025 sped up the starting code by about threefold on average; Mythos Preview in April 2026 reached about 52-fold. Experienced researchers working on the same task for four to eight hours achieved roughly fourfold. This comparison shows Claude’s ability to iterate when the goal and acceptance test are clear; the 52-fold figure is not a speedup of real-world large-model training.

Agents are also beginning to tackle more open-ended experiments. In a study of whether a weaker model can supervise a stronger one, two researchers spent about a week moving performance roughly 23% of the way from the weaker-model baseline toward an ideal upper bound. Claude agents worked about 800 hours in aggregate, at a compute cost of roughly $18,000, and closed about 97% of that gap. They proposed hypotheses, divided work, tested ideas, communicated, and iterated on their own. But humans set the research question and scoring rules, and the result did not transfer cleanly to production-scale models. The key point is that people chose the problem; AI could explore how to solve it.

What is hardest to hand over: Research judgment

Both experiments have the same precondition: people first chose the problem. Harder to delegate is recognizing which question is worth studying, deciding whether a promising result is reliable, and knowing when to abandon a direction. Anthropic calls this research taste and judgment; for now, it remains a major human advantage.

Claude has made progress on local judgments. Anthropic selected 129 points in internal research sessions where a human had temporarily gone down the wrong path. The model saw only the material available before that turn and proposed a next step. A Claude judge with access to the later outcome rated Opus 4.5’s suggestion better at 51% of those points and Mythos Preview’s at 64%. The set deliberately selected moments when humans went wrong. As a comparison, at another 127 points where the human had made a good choice, the model’s suggestion was rated better only about 20% of the time. Research does not run automatically after the topic is selected: each new experimental result brings a decision about what to investigate or change next. Claude is beginning to help people avoid wrong turns at these local forks. If that ability keeps improving, research taste may also become something AI can learn.

There is an important objection: if choosing the question determines breakthroughs, how much does it matter that AI can do the rest faster? Anthropic’s answer is that breakthrough ideas are rare, while everyday progress more often comes from scaling experiments, spotting problems, fixing them, and trying again. Even if people continue to choose the direction, compressing those repeated iterations would let each researcher advance more work at once. That is why the article argues that research could keep accelerating.

Acceleration creates new bottlenecks. Coding and experimentation consume less human execution time, but they still require compute and someone to judge the results. Code is generated faster than people can review it; teams cannot select and carry forward every experimental proposal. Amdahl’s law is straightforward here: overall speed is limited by the stages that do not speed up alongside the rest. Anthropic says both kinds of congestion have appeared internally. Employees also describe another side of the change: one says they have not personally written code for about five months; asking a colleague to run a script once created a chance to exchange ideas and help one another, but Claude now completes it quickly; another realized during a system failure that they no longer fully understood what ran every day. As output grows, human judgment and understanding become more important to preserve. What happens next depends in part on whether these bottlenecks can be resolved, and on who ultimately decides the research direction.

Three paths from here

Path one: Capability growth slows, while existing AI spreads widely. The curve might flatten because current training methods stop producing better research judgment, requiring a new technical idea. Or the supply of chips, power, grid capacity, and interconnects might fail to keep up. Even if capabilities stopped at today’s level, adoption would still change work. Anthropic points to Project Glasswing, saying Mythos Preview found more than 10,000 high-severity and critical vulnerabilities in its first weeks, shifting a bottleneck in cyber defense toward fixing them. It also imagines a 100-person company using agents to do work that once required a thousand people. Anthropic considers this path relatively less likely, but it would give society more time to adapt.

Path two: Humans choose the direction while AI continues to accelerate research. People choose questions and verify results; agents do more and more coding, experimentation, and analysis. Anthropic thinks the present evidence points most strongly to this path and imagines a 100-person organization eventually doing work that once required ten thousand or even a hundred thousand people. Public services and knowledge work could benefit, while the same capabilities could be used for mass surveillance or manipulation tailored to individuals. Code review, resource allocation, and organizational decisions would repeatedly become new speed limits.

Path three: AI builds the next generation of AI itself. If AI can also take over research direction and the design of successor systems, model development speed would depend more heavily on compute and algorithmic efficiency, while people would focus mainly on oversight and verification. Automated research might extend to drugs and robotics. Stronger models might discover new ways to align AI safely, or even stop themselves when safety cannot be assured. Another possibility is that deviation accumulates across generations of updates humans cannot understand, eventually leading to loss of control. If human labor ceases to be competitive, how the economy should be organized is also hard to predict. Even if development cycles run rapidly inside a laboratory, drug validation, grid construction, institutions, and trust between people each follow their own timescales.

These are three conditional futures, not three stages that must occur in sequence. Change the conditions below to see the branch points in Anthropic’s account: whether the capability trend continues, and who decides the research direction.

The three paths leave different amounts of time to prepare. If development continues to accelerate, finding time for safety research and institutional adjustment becomes an immediate question rather than a distant one.

Why Anthropic wants to discuss a “verifiable pause”

If AI development outruns society’s ability to adapt and conduct safety research, Anthropic wants to retain the option of coordinating a slowdown or pause in frontier development. The difficulty is getting frontier laboratories in several countries to act under the same conditions. They would first need to detect who might secretly keep training, then verify that everyone complies. Training runs are easier to conceal than traditional arms facilities, and a party that quietly continued might gain the lead. Past arms-control verification took decades to build facilities and trust; AI development may not leave that much time.

A single laboratory can stop immediately. But if other developers continue, that may mostly change who is in the lead, and could let a less cautious developer catch up without giving society more time to deliberate. Anthropic therefore favors coordinated action: if other frontier developers also slowed or paused under a verifiable arrangement, it expects it would do the same. The article calls for the capacity to make such an arrangement; it does not present an existing international verification mechanism. Anthropic wants policymakers, researchers, companies, and the public to discuss when it should be triggered, how to check compliance, and when to resume.