Artificial intelligence is beginning to play a bigger role in developing the next generation of AI. At Anthropic, that shift may already be happening faster than researchers anticipated.
A new investigation by AI research organization METR suggests that Anthropic’s AI development has accelerated to roughly twice its previous pace, with AI-powered research tools contributing significantly to the improvement.
The findings have raised concerns among AI safety researchers, particularly as increasingly capable models begin assisting with the development of their successors.
Another issue has emerged around transparency. Although Anthropic allowed independent researchers to examine parts of its internal development process, substantial portions of the supporting evidence were removed before the report became public.
The situation raises an important question: Can AI safety oversight keep pace with an industry where the technology itself is helping accelerate research?
METR Investigation Reveals Rapid Progress in Anthropic AI Development
METR spent approximately 11 weeks examining Anthropic’s internal AI research and development activities. The investigation involved access to confidential technical information, employee interviews, and model performance data that would normally remain within the company.
The researchers found evidence suggesting that Anthropic had achieved the equivalent of approximately two years of historical AI progress within a single year. This represents a notable acceleration compared with earlier development trends.
What makes the findings particularly significant is the apparent role of AI assistance. Traditionally, improvements in frontier AI models have depended heavily on computing resources, specialized researchers, and advances in training methods. METR’s assessment suggests that AI tools are becoming another major contributor to faster development.
Researchers also reported no clear evidence of immediate bottlenecks or diminishing returns that would necessarily prevent further acceleration. However, the findings do not establish that the current pace can continue indefinitely.
Anthropic Employees Report Major Productivity Gains From AI Tools
Inside Anthropic, AI systems are changing how researchers approach technical work. Employees increasingly rely on advanced models to assist with programming, research, experimentation, and other demanding engineering tasks.
According to METR’s investigation, employees using Anthropic’s most capable internal model estimated that they were approximately 12 times as productive as they would have been without AI assistance.
These figures reflect employee estimates rather than independently measured productivity improvements. Nevertheless, they indicate how important AI-powered tools have become within the company’s research environment.
The reported productivity gains were also substantially higher than estimates associated with an earlier internal model, identified as Mythos 5. This suggests that improvements in AI capabilities may translate into greater research efficiency.
Human researchers still play a central role in directing experiments, evaluating results, and making development decisions. The report does not establish that Anthropic has developed a fully autonomous system capable of independently creating its successors.
Still, the growing contribution of AI tools could allow research teams to complete increasingly complex work in less time.
Three Unreleased AI Models Highlight Anthropic’s Development Pace
METR’s investigation also provided a glimpse into Anthropic’s internal model development pipeline. Researchers examined information concerning several AI systems that the company had trained but had not released publicly.
Over approximately six months, Anthropic completed training on three models identified in the report as Model 2, Model 3, and Model 4. Their training periods extended from July through October 2026, reflecting an active research pipeline beyond the company’s commercially available systems.
The investigation also identified an older unreleased model dating back to 2022.
The existence of unreleased models is not unusual among major AI laboratories. Companies frequently train experimental systems to test technical approaches, evaluate capabilities, and investigate potential improvements before deciding whether a model is suitable for public deployment.
However, the number of models under development provides additional context for METR’s broader findings. Anthropic appears to be maintaining an intensive research schedule while integrating increasingly capable AI tools into its development process.
Whether these internal models will eventually become commercial products remains unclear.
AI Could Soon Play a More Autonomous Role in Its Own Development
One of the report’s most consequential findings concerns the possibility that AI systems could eventually make more independent contributions to AI research.
At present, much of Anthropic’s reported acceleration appears to come from AI tools assisting human employees. Researchers use these systems to handle technical assignments, support experiments, and reduce the time required for certain development activities.
METR’s investigation suggests that more autonomous research contributions could become increasingly important as AI capabilities improve.
This possibility is closely connected to the concept of recursive AI self-improvement, in which increasingly capable systems contribute to the development of more advanced versions of themselves.
Such a process could create a feedback loop. Better AI models might improve research productivity, allowing developers to build stronger models more quickly. Those systems could then provide additional assistance in subsequent development cycles.
There are important uncertainties. The report does not demonstrate that Anthropic has achieved unrestricted autonomous self-improvement or that an uncontrollable intelligence explosion is underway.
Practical constraints involving computing resources, research complexity, evaluation requirements, and technical limitations could still affect future progress.
Even so, METR’s findings suggest that AI-assisted research is already contributing to measurable changes in the speed of frontier model development.
Anthropic Faces Criticism Over Redacted Research Findings
The investigation has also generated debate over how much information leading AI companies should disclose about their internal research activities.
Anthropic voluntarily provided METR with access to sensitive technical information, allowing independent researchers to examine aspects of its development process that would otherwise remain confidential.
However, the company retained control over which details could appear in the publicly released assessment.
METR reportedly submitted an approximately 100-page evaluation, but Anthropic removed substantial portions of the supporting evidence before publication.
The research organization acknowledged that Anthropic had shared more information than initially anticipated. Nevertheless, some requests for additional access were denied, and the resulting redactions limited the amount of evidence available for independent public examination.
Anthropic co-founder Sam McCandlish defended the company’s decision, citing concerns that detailed technical disclosures could expose commercially sensitive information and help competitors accelerate their own AI development.
The disagreement highlights a difficult problem for the industry. Independent evaluations require sufficient evidence to support meaningful scrutiny, yet AI laboratories have commercial and security reasons to restrict access to certain research methods.
For outside experts, the missing information makes it harder to independently assess the report’s conclusions and determine whether alternative explanations could account for some of the observed acceleration.
AI Safety Researchers Question Whether Oversight Can Keep Pace
The METR findings have renewed concerns about whether existing AI safety practices can adapt to increasingly rapid technological development.
Frontier AI companies typically conduct evaluations to identify dangerous capabilities, assess reliability, and determine whether additional safeguards are necessary before releasing advanced systems.
Those processes become more challenging when development cycles shorten and newer models emerge before researchers have fully examined their predecessors.
AI-assisted research could increase the pressure on evaluation teams, particularly if increasingly capable systems begin performing technical tasks with less direct human involvement.
Independent oversight presents another challenge. External researchers often depend on voluntary cooperation from AI developers to access confidential models, internal testing data, and technical documentation.
The Anthropic investigation demonstrates both the value and limitations of that arrangement. METR gained unusual access to internal information, but restrictions on publication prevented the organization from sharing much of its supporting evidence.
Faster AI development does not necessarily mean that safety standards are deteriorating. Companies may also use advanced AI tools to improve testing, identify vulnerabilities, and strengthen safeguards.
However, the report reinforces the importance of ensuring that safety research progresses alongside improvements in model capabilities.
What Anthropic’s Accelerating AI Research Means for the Industry
Anthropic’s experience may offer an early indication of how AI development could change across the broader technology industry.
For years, competition among leading AI companies has centered on computing infrastructure, research talent, training data, and access to capital. Those factors remain important, but the ability to use AI systems effectively within research operations could become another major competitive advantage.
Companies that successfully integrate advanced models into engineering workflows may be able to shorten development timelines, conduct experiments more efficiently, and explore technical approaches that previously required substantial human effort.
The implications extend beyond commercial competition. Faster AI research could support scientific discovery, software development, and new enterprise applications. It could also create additional pressure on governments and independent organizations responsible for assessing emerging risks.
Much remains uncertain about how long the current acceleration can continue. Technical bottlenecks may emerge, productivity gains could level off, and increasingly complex models may require more extensive evaluation.
The METR investigation also leaves unanswered questions because important supporting evidence remains confidential.
What the findings do suggest is that AI is becoming more than the outcome of technological research. It is increasingly becoming part of the process that drives further progress.
For Anthropic and the wider AI industry, that shift could reshape how quickly advanced systems reach the market.
The challenge now is ensuring that the ability to evaluate, understand, and govern those systems develops quickly enough to keep pace.
