Anthropic Warns It Is Very Close to Recursive Self-Improvement, the Next AI Milestone

Anthropic has warned that artificial intelligence could soon reach a stage where it significantly accelerates the development of more advanced AI systems. The company disclosed the possibility in an amended IPO filing submitted to the US Securities and Exchange Commission (SEC) on November 2, 2026. The filing says Anthropic’s internal estimates suggest future models may imminently meet its definition of dramatically accelerated progress attributable to AI, a capability associated with recursive self-improvement (RSI).

The disclosure has attracted attention because RSI could change how quickly AI technology develops. Instead of relying almost entirely on human researchers to design, train and improve models, increasingly capable AI systems could perform a growing share of that work themselves. However, Anthropic has made an important distinction: it has not designated any current model as having crossed its RSI-related threshold, and it does not expect to safely allow Claude to build successive generations of frontier AI models fully autonomously without meaningful human supervision.

Key Takeaways

• Anthropic warned in its November 2 IPO filing that it may imminently reach an internal threshold for dramatically accelerated AI-driven research progress.
• The warning concerns future models, not a confirmed capability of current Claude models.
• Recursive self-improvement describes AI systems helping accelerate the development of more capable AI systems.
• Anthropic says Claude already contributes to model research, training data generation, evaluations, safety audits and software development.
• More than 90% of Anthropic’s code contributions in September 2026 were written by Claude, according to the filing.
• The company says the value of its internal Claude usage at list prices increased more than 250 times between January and September.
• Anthropic is considering changes to its safety policy to maintain human oversight of frontier model development.

What Is Recursive Self-Improvement (RSI)?

Recursive self-improvement is a process in which an AI system helps improve the technology used to build more capable AI systems.

In a conventional AI development process, human researchers make important decisions about model architecture, training methods, datasets, computing resources and evaluation procedures.

AI tools can already assist with many of these tasks. They can write code, analyse research papers, generate training data, identify problems in experiments and automate parts of model evaluation.

RSI becomes more consequential when these contributions significantly accelerate the development of subsequent AI generations.

The process could create a feedback loop:

1. Researchers develop an AI model.
2. The model helps researchers improve training methods and software.
3. Those improvements help create a more capable model.
4. The newer model becomes more effective at AI research and development.
5. The cycle repeats, potentially accelerating progress.

This does not automatically mean an AI system becomes conscious or independently decides to build itself. RSI is primarily about the ability of AI systems to accelerate their own technological development.

What Did Anthropic Disclose in Its IPO Filing?

The most important statement appears in the risk factors section of Anthropic’s amended Form S-1 filing.

The company said its internal estimates and expectations about future models indicate that it may imminently meet its Responsible Scaling Policy’s definition of “dramatically accelerated progress attributable to AI.”

However, Anthropic also explicitly stated that it has not designated any current model as exceeding this threshold and continues to assess the underlying data.

The disclosure also identifies a measurement problem: much of the evidence used to estimate AI-driven acceleration comes from lagging indicators. This means researchers may recognise that development has accelerated only after the acceleration has already happened.

That creates a potential challenge for AI safety teams. If model capabilities improve faster than expected, safety evaluations, security systems and alignment techniques may struggle to keep pace.

Claude Is Already Helping Anthropic Build Better AI Models

One reason the disclosure matters is that Anthropic is already using Claude extensively in its own research and development operations.

The company’s IPO filing describes a process in which each new version of Claude helps accelerate the development of the next.

Claude is involved in several areas of internal development:

• Running research experiments.
• Generating and processing training data.
• Automating model evaluations.
• Supporting safety audits.
• Writing and reviewing software code.
• Helping researchers investigate technical problems.

Anthropic also disclosed that more than 90% of contributions to its codebase in September 2026 were written by Claude.

The company said the value of its internal Claude usage, calculated using customer list prices, increased more than 250 times between January and September 2026.

In April, the monthly value of that usage first exceeded monthly personnel expenses, excluding stock-based compensation.

This does not mean Anthropic was spending more actual money on Claude than on employee salaries. The comparison uses the list-price value of internal usage rather than the company’s actual internal costs.

Nevertheless, it illustrates how deeply AI tools have become integrated into the company’s engineering and research operations.

Why RSI Could Accelerate AI Development

The potential importance of RSI comes from the possibility of a compounding improvement cycle.

AI research involves several time-consuming activities, including writing code, testing hypotheses, analysing results and improving training infrastructure.

If AI systems become significantly better at these activities, researchers could potentially conduct more experiments in less time.

That could produce several effects:

Faster model development: AI systems could help researchers identify more effective training techniques and improve development workflows.

More efficient computing: AI-assisted research could identify ways to use computing resources more effectively, allowing companies to train and operate more capable models.

Improved scientific research: More capable systems could assist with mathematics, engineering, materials science and other research fields.

Shorter development cycles: If AI systems become more effective at research and experimentation, the time required to develop subsequent generations of models could decrease.

However, these outcomes are possibilities rather than guaranteed consequences. The speed and scale of any acceleration will depend on technical progress, available computing resources, research quality and the effectiveness of human oversight.

Anthropic Says Human Oversight Will Remain Important

Despite warning about the possibility of accelerated AI research, Anthropic says it does not anticipate safely using Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding.

The company is also considering updates to its Responsible Scaling Policy (RSP) to keep humans involved in decisions about frontier model development.

The policy establishes safety requirements connected to different levels of AI capability and risk.

As models become more capable, Anthropic evaluates whether its safety, alignment and security measures remain adequate.

The company says it may delay training, restrict access or postpone model releases when necessary.

It has also disclosed that it previously delayed pre-training a new frontier model for more than a month to strengthen security measures designed to protect model weights and intellectual property.

The latest filing indicates that Anthropic may introduce additional human review and approval requirements as AI systems perform a larger share of its research and development work.

Why AI Safety Researchers Are Concerned

The main concern is that AI capabilities could advance faster than researchers can understand and control them.

If AI systems become increasingly capable of designing experiments, modifying software and improving model training, developers may find it harder to predict how quickly their technology will advance.

Anthropic specifically warns that its ability to identify acceleration is limited and that safety, alignment and security measures may not develop at the same speed as model capabilities.

Potential risks include:

• Reduced human control: Increasingly autonomous systems could make important technical decisions with less direct human involvement.
• Safety gaps: Evaluation and monitoring systems could fall behind rapidly improving capabilities.
• Security threats: More capable systems could increase the potential for cyber misuse.
• Unanticipated behaviour: Developers may struggle to predict how complex systems behave when given more autonomy.
• Competitive pressure: Companies could face incentives to move faster even when additional safety testing is needed.

These risks do not establish that an AI system will inevitably become uncontrollable. They explain why developers are trying to establish safeguards before more advanced capabilities emerge.

What This Means for the AI Industry

Anthropic’s disclosure could influence how the broader AI industry discusses model development and safety.

The company is effectively warning investors that AI is becoming an increasingly important part of the process used to build AI itself.

This could have implications for competing laboratories, governments, regulators and companies investing in AI infrastructure.

If AI-assisted research produces significant productivity gains, companies with access to advanced models and large computing resources could potentially increase their development pace.

At the same time, the ability to maintain reliable safety testing and human oversight could become an increasingly important competitive and regulatory issue.

Anthropic has also publicly called for “pacing the frontier”, arguing that AI capabilities should not advance faster than the safety and security measures needed to support them.

The Bigger Picture

For years, recursive self-improvement has been discussed as a possible turning point in artificial intelligence. Anthropic’s latest filing brings the subject into a formal corporate disclosure, making it relevant not only to AI researchers but also to investors and policymakers.

The most significant detail is not that AI has already started independently building its successors. It is that Anthropic says its future models may soon meet an internal threshold for dramatically accelerated AI-driven research.

The company is already using Claude extensively to support its own development process. If that contribution grows substantially, the relationship between human researchers and AI systems could change significantly.

Looking Ahead

The next important development will be Anthropic’s assessment of whether future Claude models actually cross its Responsible Scaling Policy threshold. The company has said it continues to evaluate the underlying data and plans to provide further reports. Any changes to its safety policy could also establish how much human approval will be required as AI systems take on more research responsibilities.

For the wider industry, the central question is whether AI can help accelerate its own development without making safety, security and human oversight fall behind. Recursive self-improvement could create substantial scientific and economic opportunities, but its consequences will depend on how reliably companies manage the technology as it becomes more capable.