Anthropic has disclosed that artificial intelligence is increasingly automating its own development, with Claude writing over 80% of the code in the company's repositories as of May 2026. This dramatic shift, which occurred in just 16 months, signals a fundamental change in how AI systems are built and raises urgent questions about whether safeguards can keep pace with the technology's acceleration.

How Is AI Building Itself Faster Than Humans Can?

On June 4, 2026, Anthropic's internal research arm published a report titled "When AI builds itself," detailing a phenomenon called recursive self-improvement. The concept is straightforward but unsettling: AI designs its own next-generation successor, which then creates the one after that. With each cycle, development speed increases.

The numbers are striking. When Claude Code was first released for research in February 2025, the AI contributed only single-digit percentages of code to Anthropic's repositories. By May 2026, that figure had jumped to over 80%, with Anthropic estimating it could exceed 90% if scripts and experimental code are included. The company emphasized that 80% represents a conservative measurement focused on core code quality.

Productivity metrics have surged alongside this shift. In the second quarter of 2026, the amount of code contributed by a single engineer was eight times higher than it was four years ago. However, Anthropic acknowledged that lines of code measure quantity, not quality, and the eightfold figure likely inflates actual productivity gains. Still, the company confirmed the underlying trend: more work is being delegated to Claude.

What Tasks Can AI Now Handle Independently?

Perhaps more concerning than code volume is the duration of tasks AI can complete without human intervention. According to measurements by external evaluation agency METR, the time AI spends handling tasks independently is doubling roughly every four months.

The progression is dramatic:

• March 2024: Claude Opus 3 could independently perform a task that took a human 4 minutes
• March 2025: Claude Sonnet 3.7 handled a task requiring 1 hour and 30 minutes of human work
• March 2026: Claude Opus 4.6 processed a 12-hour task on its own

Anthropic stated that if this trend continues, the company will move from delegating short errands to handing over entire projects that take several days. This escalation raises fundamental questions about human oversight and control.

The company divides AI development into two domains: execution (writing code and running training) and judgment (deciding what to attempt and interpreting results). Humans have traditionally held the advantage in judgment, relying on research intuition to discern what is worth creating. However, even this final frontier is showing signs of erosion.

Is AI Judgment Becoming Better Than Human Intuition?

In execution tasks, AI has already surpassed human performance dramatically. When Anthropic asks new models to optimize code for speed while keeping results identical, the improvements have accelerated sharply. In May 2025, Claude Opus 4 made code run about 3 times faster. By April 2026, Claude Mythos Preview achieved a 52-fold speedup. A skilled human typically spends four to five hours to achieve a 4x improvement, making AI's progress remarkable.

The judgment domain presents a more unsettling picture. In April 2026, Anthropic released results from an experiment where Claude-based AI handled an entire safety research task independently. The AI formed hypotheses, ran experiments, reported results, and decided on next steps. A problem that took two human researchers a week to reach 23% completion was solved to 97% by several AIs using a cumulative 800 hours and approximately $18,000 worth of computing resources.

Another experiment tested AI's intuition for choosing the right direction when research hits a dead end. Anthropic selected 129 moments where human researchers had actually gone down the wrong path, showed the AI only the situation immediately preceding those moments, and asked for the next move. The rate at which AI found a better method than the human rose from 51% with Claude Opus 4.5 in November 2025 to 64% with Claude Mythos Preview six months later.

"Humans have the edge in judgment, but even that is wavering," Anthropic noted in the report, acknowledging that while humans still chose the problems and set grading criteria, the results signal an early warning that AI is making increasingly better choices at research crossroads.

Anthropic Institute, Research Report

Code quality expectations have also shifted. Within Anthropic, the prevailing view is that while Claude-written code was inferior to human work until the end of 2025, it is now on par with human quality and will surpass it within a year.

What Three Futures Does Anthropic Envision?

Anthropic outlined three possible futures as AI self-improvement accelerates:

• Growth Plateau: The current steep upward curve breaks at some point. Anthropic considers this unlikely, noting that none of the capabilities measured so far have shown signs of plateauing
• Automated Development with Human Control: AI development becomes fully automated, but humans retain the steering wheel. This is the scenario Anthropic considers most likely, creating a world where a 100-person company could do the work of an organization of tens of thousands
• Full Autonomy: AI finally designs and creates its own successors with humans only handling supervision and verification. If AI heads in an unintended direction, humans cannot correct it, making alignment and safety crucial

The second scenario carries its own risks. The same power that enables efficient AI development could be mobilized to monitor entire populations or spread personalized public opinion manipulation. The third scenario is even more precarious, as rare misaligned behaviors could be passed down from successor to successor, growing incrementally until they reach a point beyond human control.

What Is Anthropic's Proposed Solution?

Anthropic argues that any company building AI must have mechanisms in place to slow down or pause development when necessary. However, the company acknowledges a critical problem: if only one company stops, competitors could pull ahead. Therefore, Anthropic emphasizes that the real key is for the entire world to pause together in a way that allows everyone to verify that others have truly stopped.

The company notes that creating a collective "brake" is harder than controlling nuclear weapons. Unlike nuclear development, AI advancement is distributed across multiple private companies, academic institutions, and countries, making coordinated action extraordinarily difficult. Anthropic's warning essentially amounts to a call for unprecedented global cooperation on technology development.

How Are AI Leaders Addressing Biosecurity Risks?

Beyond internal development concerns, Anthropic and other AI leaders are pushing for regulatory action on biosecurity. In an unprecedented show of unity, leaders from OpenAI, Anthropic, and Microsoft co-signed an open letter to Congress urging immediate legislative action to close what they call a "dangerous biosecurity gap." The signatories include Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman, and Microsoft AI CEO Mustafa Suleyman.

The core concern is that AI models like Claude 3.5 Sonnet could theoretically assist malicious actors in designing biological threats by providing high-level reasoning and synthesis of complex biological data. The primary focus of the open letter is regulating synthetic DNA and RNA providers. Currently, anyone can order genetic material online, which can then be assembled in a laboratory setting.

While many reputable companies voluntarily screen orders for known pathogens, the practice is not universally mandated by law. AI leaders are calling for a federal requirement that all companies selling synthetic genetic material must screen purchases for sequences associated with dangerous biological agents. As AI models become more adept at protein folding and genetic engineering, the risk of jailbreaking a model to generate a lethal sequence increases.

The move by Altman, Amodei, and Suleyman is a strategic attempt to preempt heavy-handed regulation by proposing specific, targeted rules. By focusing on the physical side of the problem, DNA synthesis regulation, they aim to protect the digital freedom of AI development. If the industry can prove it is proactive about safety, it may avoid broader, more restrictive legislation that could stifle innovation.

Anthropic's disclosure of its internal data represents a significant shift in transparency for the company. By revealing that Claude now writes over 80% of its own code and demonstrating the accelerating timeline for AI task autonomy, the company is essentially sounding an alarm: the pace of AI self-improvement is outpacing institutional readiness to manage it. Whether global coordination on AI development can be achieved remains an open question, but Anthropic's report makes clear that the window for action may be closing faster than previously anticipated.