Anthropic published something today that is rare for a frontier lab. Not a model, not a product, but a direct, named warning about its own trajectory. In a piece titled "When AI builds itself," authored by Jack Clark and Marina Favaro of the Anthropic Institute, the company laid out internal data showing that Claude is now meaningfully accelerating the development of Claude. Their own framing, posted from the official Anthropic account: their data shows Claude is accelerating AI development, a possible path to recursive self-improvement, or AI autonomously building a more capable successor, and "it's happening faster than we thought." The reason this matters is the source. This is not a hype thread from an anonymous account or a benchmark blog chasing traffic. This is the company that builds Claude putting its name on the claim that AI is starting to build AI, and arguing the implications deserve far more attention than they are getting. Here is what the data actually shows, what it does not show, and why your reading of it should be neither dismissive nor breathless.

Anthropic published internal numbers it had never released before. These are the ones that matter.

80% of their code is now written by Claude. As of May 2026, more than 80% of the code merged into Anthropic's production codebase was authored by Claude. Before Claude Code launched in research preview in February 2025, that number was in the low single digits. Leadership has publicly estimated the figure is 90% or higher when you include scripts and experimental code.

Engineers ship 8x more code per quarter. For Anthropic's first four years, code merged per engineer stayed flat. It began climbing in 2025 when Claude started running code instead of just suggesting it, then steepened again in 2026 as models began working autonomously over longer horizons. The typical engineer now merges roughly 8 times as much code per day as in 2024. Anthropic is honest that lines of code overstates true productivity, but the inflection is real.

The training-speedup test went from 3x to 52x in a year. Every time Anthropic releases a model, it runs the same test: give Claude code that trains a small AI model, and ask it to make that code run as fast as possible. A skilled human researcher needs four to eight hours to hit a 4x speedup. In May 2025, Claude Opus 4 averaged about 3x. By April 2026, Mythos Preview averaged about 52x. In that one slice of research work, Claude went from helpful to superhuman in under a year.

Claude is starting to beat humans on research judgment. Anthropic took real research sessions where a human took a wrong turn, showed Claude only the work up to that point, and asked what to do next. A separate Claude that could see how things actually turned out judged whose call was better. In November 2025, the best model beat the human choice 51% of the time. By April 2026, Mythos Preview hit 64%. The caveat Anthropic states plainly: these were deliberately chosen moments where the human's choice had room to improve, so it is not a clean head-to-head. But the direction is the point.

One case in numbers: in April 2026, Claude shipped over 800 fixes that reduced a class of API errors by a factor of a thousand. The engineer overseeing it estimated a human would have needed four years.asswing partners, plus Anthropic's own teams, have now found more than 10,000 high or critical-severity vulnerabilities in essential software. Anthropic's public coordinated disclosure dashboard, updated May 22, lists 1,596 specific disclosures across 281 open-source projects with 97 patched so far. This is not a chatbot. It is a vulnerability-finding engine that has already shifted the global security baseline.

That history is exactly why the Mythos 1 sighting matters. The model is leaving the controlled environment.

So if Claude is already building Claude, what is left for the humans? Precisely one thing, and it is the thing worth understanding.

Anthropic is careful to say what today's data does not show: that Claude can choose what to build. Claude can take an underspecified engineering problem and figure out the method. It can match or beat skilled humans at running a well-defined experiment. What it cannot reliably do yet is exercise research taste, deciding which problems are worth working on, which results to trust, and when an approach is a dead end.

Put simply, Claude is doing the building. Humans are still doing the directing. Edison said genius is 1% inspiration and 99% perspiration. Anthropic's data shows the 99% perspiration getting automated fast, while the 1% inspiration stays human. The unsettling part is the line they add next: research taste might just be another capability AI fails at for a while and then learns, the way it eventually learned to explain why a joke is funny. They do not know if that gap closes. They are saying it might, sooner than most institutions are prepared for.

This is the question worth sitting with, because a company about to file the largest AI IPO in history does not usually publish a document about possibly losing control of its own technology.

Two honest readings, both probably true at once. The first is mission. Anthropic was founded on AI safety, and publishing internal evidence that AI is accelerating AI development, alongside a call for the ability to slow down or pause if needed, is consistent with that founding purpose. The piece ends with a genuinely sober proposal: that the world should build the capacity to verify whether frontier labs have actually slowed or paused, so that a coordinated slowdown would not just hand the lead to whoever cheats. That is not marketing. That is a company arguing against its own short-term incentive to race.

The second reading is positioning. "Our model is so capable it is starting to build its successor" is, whatever else it is, the single most powerful enterprise sales and recruiting message a frontier lab could send the week it picks its IPO bankers. Both things can be true. The data is real, the safety concern is real, and it also happens to be the strongest possible statement of capability at the most valuable possible moment. Your high-ticket readers can hold both of those in their heads, and they will trust you more for naming both than for picking one.

If you are an engineer, the role Anthropic describes is already arriving at your job: the value moves from writing code to specifying and reviewing it. Anthropic flat-out says once AI and human code quality reach parity, which they expect within a year, humans stop writing code and shift to reviewing it, and the new bottleneck becomes how fast humans can review what the AI produces. Position yourself for the judgment layer, not the typing layer.

If you run a team or company, the line to internalize is that a 100-person company is starting to be able to do the work of a 1,000-person one, because each person sits atop a stack of agents. The competitive question for the next two years is not whether you adopt AI, it is whether you reorganize around it faster than your competitors do.

If you invest or watch the frontier, this reframes the entire AI race. If the labs that best automate their own research compound fastest, then the lead goes to whoever has the best models and the most compute to run them in a loop, which is exactly the position Anthropic, OpenAI, and Google are spending hundreds of billions to secure. The Karpathy hire we covered last month, "use Claude to train the next Claude," was not a one-off. It was Anthropic telling you this exact strategy out loud, two weeks before showing the data behind it.

Anthropic published "When AI builds itself" today, releasing internal data showing Claude is now accelerating Claude's own development, what it calls a possible early path to recursive self-improvement. The numbers: over 80% of Anthropic's merged code is now written by Claude, up from low single digits before Claude Code launched in early 2025. Engineers ship 8x more code per quarter than in 2024. On the recurring training-speedup test, Claude went from a 3x improvement in May 2025 to 52x with Mythos Preview by April 2026, against a skilled human's 4x. On research next-step judgment, Mythos beat the human choice 64% of the time, up from 22% in 2024. The honest limit: Anthropic says AI still lacks research taste, knowing which problems matter, and that gap is the only thing standing between today and a system that designs its own successor. They are not claiming that future is here. They are warning it might come faster than the world is ready for, and calling for the ability to verifiably slow down if it does. The most important sentence is the quiet one: the part of the work that is still human is shrinking, and even taste might just be the next capability AI learns. Claude is helping build Claude. The only question left is how much of the judgment stays ours.

Stay building. 🤖