Anthropic Claims the Path to AI Self-Improvement Is Becoming Clear, Sparking Debate Over IPO Narrative

By LetterAI

Yesterday, Anthropic published a long article titled "When AI Builds Itself," which sounds like something from an Isaac Asimov science fiction novel, and the theme is indeed a highly sci-fi concept: recursive self-improvement.

In simple terms, previously, human researchers wrote code, ran experiments, and trained models to make AI stronger. But if AI itself begins to participate in designing, training, testing, and optimizing its own successor versions, the pace of AI advancement will no longer be driven solely by humans—it may begin to be fueled by AI’s “self-evolution.”

To this end, Anthropic has issued a call:

We believe it would be greatly beneficial for the world if it could choose to slow down or temporarily halt the development of frontier AI, allowing societal structures and alignment research to catch up with technological progress.

This sounds like a security warning, but given the timing of Anthropic’s impending IPO, it’s hard not to interpret it as another narrative rehearsal: Claude is so useful that it’s already beginning to create the next generation of Claude.

To illustrate that AI is increasingly involved in AI research itself, Anthropic presented a large amount of internal data.

For example, as of May 2026, over 80% of the code merged into Anthropic’s codebase was written by Claude. Before the release of Claude Code, this figure was only in the single digits.

By the second quarter of 2026, according to Anthropic’s statistics, the amount of code merged daily by engineers had increased by approximately eight times compared to 2024.

More noteworthy than the amount of code is that Claude is tackling more open-ended engineering problems.

Anthropic states that over the past year, the frequency with which employees have had to correct Claude, steer it back on track, or take over tasks midway has been consistently decreasing. This trend is not limited to simple tasks but also applies to the most complex open-ended tasks.

An open task refers to an issue without a clear set of instructions—for example, a system crash or a failed training job—where engineers initially don’t know what the solution looks like and must troubleshoot while analyzing the problem.

In the past, these tasks relied most heavily on human experience, but in the most open-ended tasks, Claude’s success rate reached 76% by May 2026, a 50-percentage-point increase within six months.

Beyond writing code, Anthropic also uses Claude for code reviews, identifying bugs, security vulnerabilities, and other defects. Their retrospective analysis found that if Claude had automatically reviewed every code change in the past, about one-third of the bugs that caused outages on claude.ai could have been caught before deployment.

Furthermore, Claude has begun participating in the research process.

Anthropic has a fixed test: given Claude a piece of code for training a small model, it must find ways to make the code run faster without altering the output results. In May 2025, Claude Opus 4 achieved approximately a 3x speedup; by April 2026, Claude Mythos Preview had pushed this figure to about 52x.

Anthropic also mentioned an open AI safety research case: they posed a question to a Claude-powered agent—can a weaker model reliably supervise a stronger one?

This process requires formulating hypotheses, testing them, sharing findings with parallel agents, and iterating repeatedly.

Two human researchers spent a week closing about 23% of the gap, while Claude closed 97% after accumulating approximately 800 hours of computation and $18,000 in compute costs.

Of course, this result has limitations—the problems were selected by humans, the evaluation criteria were defined by humans, and the results have not yet been fully scaled to production-grade models. But it still demonstrates that Claude can now design experiments, execute them, and iterate on them independently within a research framework where humans have set the direction.

Additionally, when human researchers go down the wrong path, Claude can still provide better guidance on the next steps.

Anthropic selected 129 internal Claude Code research sessions in which human researchers collaborated with Claude to solve open-ended research problems. Anthropic identified certain points in these sessions where human researchers later turned out to have taken a detour, then provided the context up to those points to different versions of Claude to see what next steps the model would suggest. Another Claude judge, aware of the full session outcome, evaluated whether the model’s suggestion or the human’s actual choice was better.

The results show that Claude is increasingly able to suggest better next steps at nodes where human researchers have since been shown to have room for improvement.

In the past, advancements in AI models were primarily driven by human researchers and engineers. Humans decided which experiments to conduct, wrote the code, trained the models, and drove the functional evolution of AI.

Now, an increasing number of steps in this chain are being taken over by Claude.

Anthropic provided a very intuitive phase chart:

From 2021 to 2023, Anthropic was no different from ordinary tech companies—humans wrote code and documentation on laptops.

From 2023 to 2025, chatbots began integrating into workflows, with engineers having models generate code snippets that they then copied into their editors.

Between 2025 and 2026, programming agents emerged, and Claude began to autonomously write and modify code, sometimes even completing entire files independently.

Today, agents can execute code on their own and delegate tasks lasting several hours to other agents.

Next comes the stage that Anthropic truly fears: the closed loop.

If this day comes, future versions of Claude could be continuously improved by Claude itself—this is known as recursive self-improvement.

Anthropic speaks cautiously in the text: we have not yet reached that point, and recursive self-improvement is not inevitable. But it still emphasizes that the path toward it is beginning to become visible.

That’s why Anthropic concludes the article by discussing slowing down, or even pausing. It does not mean that all AI companies should shut down immediately, but rather that if the risks of AI self-improvement continue to rise, leading laboratories need a coordinated, verifiable mechanism for slowing down.

In other words, the singularity is approaching, and humanity must take control.

On the surface, this is a highly forward-thinking security document. Anthropic discusses recursive self-improvement, the possibility of AI improving itself at an increasingly rapid pace, and the need for human society to prepare ahead of time for slowdown and pause mechanisms.

But given the timing of Anthropic’s upcoming IPO, this article takes on another layer of meaning.

In a way, Factory A’s recent moves are like that annoying top student in class—they’re genuinely capable, but they’re also putting on a show.

It’s not just saying “We have a strong Claude”—it’s going a step further, saying “Claude is helping us build an even stronger Claude.”

If Anthropic were merely selling a model or a tool, it would be difficult to escape direct comparisons: Anthropic has Claude, while OpenAI has GPT; Anthropic has Claude Code, while OpenAI has Codex; Anthropic targets enterprise customers, and so does OpenAI. The competition between the two companies is extremely tight, hinging on which one can tell a more compelling story to the market.

It should be noted that just three days ago, OpenAI wrote in a document on advanced AI governance:

We are also seeing early signs of recursive self-improvement in today’s systems: the development of AI itself is being accelerated by AI.

This will intensify competitive pressures between developers and nations, creating governance challenges that existing institutions cannot address.

Three days later, Anthropic followed up by saying: The path to recursive self-improvement for Claude has begun to become visible.

If Claude develops as envisioned, it won't just be a typical product narrative—it will become a research flywheel.

Claude writes code, runs experiments, and optimizes training workflows, which in turn reduces incidents in Anthropic’s own products... Once this system is up and running, Claude will no longer be just a product of Anthropic—it will also be a vital production tool for Anthropic.

Users see Claude as a product; enterprise customers buy Claude’s capabilities—but what Anthropic truly wants the capital markets to notice is that Claude has been integrated into the foundational processes of cutting-edge model development and is now inside Anthropic’s engine room.

Capital markets love to hear about flywheel stories—like a treasure pot that continuously generates wealth: a more powerful Claude enables Anthropic’s engineers to merge more code; more code allows faster iteration of products and infrastructure; faster iteration lets researchers run more experiments; and more experiments, in turn, help make the next generation of Claude even stronger. Once the next generation of Claude becomes stronger, it further accelerates Anthropic’s R&D.

Claude’s iteration speed also fuels this flywheel. From public release dates, major model updates for Claude occurred roughly every three to four months between 2023 and early 2025, but after the introduction of Claude 4, Anthropic’s model updates have become significantly more frequent.

Claude 4 was released in May 2025, Opus 4.1 in August, Sonnet 4.5 in September, Haiku 4.5 in October, and Opus 4.5 in November.

By 2026, Opus 4.6 was released on February 5, Sonnet 4.6 on February 17, Opus 4.7 on April 15, and Opus 4.8 on May 28. The gap between Opus 4.7 and Opus 4.8 was only 42 days.

Anthropic is表面上 saying, "This could be dangerous; we need to prepare the brakes in advance," but it is also implying: "We've already seen what happens when the accelerator is pressed."

The subtlety of the IPO narrative lies herein: it simultaneously emphasizes the risks while elevating its technological standing.

Not all AI companies are qualified to discuss recursive self-improvement; you must first convince the outside world that your AI has entered the AI development process before you can claim that global coordination may be necessary.

As mentioned earlier, just before Anthropic published this long article, OpenAI had just brought recursive self-improvement to the table.

But the two companies' statements are very different.

OpenAI's paper "Democratic Governance of Frontier AI" is a policy blueprint addressed to Washington, focusing not on "how models become more powerful," but on how to govern frontier AI as it continues to advance.

Most of the content mentioned in that report is not suitable for elaboration, except for one key sentence: OpenAI says that early signs of recursive self-improvement are already visible in today’s systems.

This statement and Anthropic’s long article actually point in the same direction.

While OpenAI talks about systems, Anthropic talks about itself.

OpenAI means that AI is developing too quickly for existing governance structures to keep up, so a new set of rules is needed.

Anthropic directly revealed that system to the market, indicating that Claude has entered our development pipeline, and thus we have observed the path of AI self-acceleration.

This move is brilliantly executed—it feels like there’s already infighting inside OpenAI—this is outright theft of our idea! We were here first!

Just kidding, but OpenAI really needs to step up—hurry up and launch GPT 5.6!