AI Is Dramatically Accelerating AI Research: Inside the METR-Anthropic Report

In a development that could redefine the trajectory of artificial intelligence, a new independent report from METR (Model Evaluation & Threat Research) has concluded that AI is now dramatically accelerating AI research at Anthropic. The findings, based on unprecedented access to the company's internal R&D processes, suggest that the pace of AI capabilities progress has roughly doubled over the past year, with AI itself being the primary driver. This milestone, which Anthropic's own Responsible Scaling Policy (RSP) flagged as a critical threshold, has profound implications for AI timelines, safety, and the industry's ability to manage what could be a runaway feedback loop.

TL;DR: The Key Takeaways

Doubling the Pace: Between October 2025 and October 2026, Anthropic's AI capabilities progress accelerated to roughly twice the rate of the previous year, according to an internal version of the Epoch Capabilities Index (AECI).
AI as the Driver: METR attributes approximately 80% of the current acceleration to AI automation, a figure that is growing over time.
Threshold Crossed: This acceleration appears to meet or exceed the 'dramatically accelerated progress due to AI' threshold defined in Anthropic's RSP, which triggers enhanced safety and security requirements.
Compounding Risk: METR found no clear bottlenecks or steep diminishing returns, raising the possibility of compounding acceleration that could lead to extreme superintelligence within a year.
Industry Unprepared: The report and commentators argue that the AI industry lacks the robust security, monitoring, and alignment infrastructure needed to safely navigate this rapid acceleration.

Unprecedented Access: What METR Found

From August 12 to October 29, 2026, METR staff conducted a high-access evaluation at Anthropic, gaining unprecedented visibility into the company's AI R&D pipeline. This included metrics on pretraining, reinforcement learning, training compute, benchmark scores, AI speedup analyses, and interviews with staff. The resulting report, published on November 5, 2026, offers a rare, data-driven look at how AI is transforming the very process of AI development.

The central finding is stark: capabilities progress at Anthropic has been going about twice as fast as it had been over the previous twelve months. This measurement, based on an internal version of the Epoch Capabilities Index called AECI, is supported by a wide range of additional evidence, though much of it has been redacted for security reasons. METR notes that this redacted evidence constitutes "a majority of the evidence [they] found most informative."

Attribution: AI as the Primary Accelerant

While the initial acceleration in late 2025 was likely influenced by non-AI factors, METR's analysis indicates that AI automation is now the dominant force. The report estimates that approximately 80% of the current acceleration is due to AI, a judgment made with substantial uncertainty but supported by multiple lines of evidence. Anthropic staff self-report dramatic uplift from AI tools, with some describing a ~12x productivity boost. This means researchers are effectively experiencing ten days of work for every calendar day, thanks to AI's ability to accelerate thinking, reading, coding, and experimentation.

This attribution is critical. Anthropic's RSP defines the "dramatic acceleration" threshold as requiring that the doubling of progress be "plausibly attributable to AI automating AI R&D, rather than acceleration in compute or other inputs." METR's findings, combined with Anthropic's own statements that it "may imminently meet" this threshold, suggest that the condition has indeed been met.

continue reading below...

The Math Behind the Acceleration

To understand the magnitude of this shift, it helps to unpack the math. Overall AI capabilities can be thought of as the product of compute multipliers (algorithmic and architectural innovations) and training compute. If overall capabilities are advancing twice as fast, and this is due to software progress, then the rate of growth in compute multipliers must be even higher—perhaps 2.5-3x faster, based on reasonable assumptions.

Furthermore, since researcher labor is only one input into discovering these multipliers, the uplift to labor must be the square of the acceleration in compute multipliers. This implies a required labor speedup of roughly 4-9x, but given the need to overcome diminishing returns, the actual uplift needed today is likely around 10x. This aligns with the ~12x self-reported by Anthropic staff.

This is not merely a matter of adding more researchers; it's as if each researcher is in a time warp, able to accomplish in one day what previously took ten. This level of acceleration would be impossible through hiring alone, highlighting the transformative power of AI in the research process.

Compounding Acceleration: The Path to Superintelligence?

The most alarming implication is the potential for compounding acceleration. If AI capabilities continue to improve, the uplift to researchers will grow, leading to even faster progress. METR's report states that they "did not identify any clear bottlenecks or steep diminishing returns to further scaling that would get in the way of such compounding acceleration."

This raises the specter of a runaway feedback loop. Ajeya Cotra, a METR researcher, notes in a commentary that if this acceleration compounds smoothly, "we would reach extreme superintelligence within a year." The logic is simple: if a year of progress got us from modest acceleration to doubling, then the next doubling could take six months, then three, then six weeks, and so on. Within a year, the cumulative progress would be astronomical.

While there are theoretical limits to software efficiency, these are likely far from current capabilities. The report's authors are careful to note significant uncertainty, but the possibility of an intelligence explosion is now a concrete concern, not a distant hypothetical.

Industry Unprepared: The Safety Gap

The report underscores that the AI industry is not prepared for this scenario. Anthropic's own RSP, in its May 2025 version, required that by the time it reached the AI R&D-5 threshold (a somewhat weaker milestone), it would need security sufficient to "protect against model-weight theft by state-level adversaries." Yet recent security breaches at both OpenAI and Anthropic by amateur groups highlight the gap between current practices and what would be needed.

Beyond model weight security, experts argue that a safe intelligence explosion would require robust logging and monitoring, reliable measurements of monitorability, effective countermeasures against monitor collusion, principled alignment training, and governance infrastructure for pacing progress. None of these are in place at the required scale.

Cotra emphasizes that "the industry is nowhere near this bar." The redacted nature of much of the evidence also raises concerns about transparency, as external scientists cannot fully scrutinize the findings. This lack of transparency could hinder the collective understanding needed to manage the risks.

Broader Implications: AI Acceleration in Science

The acceleration at Anthropic is part of a broader trend of AI transforming scientific research. As noted by the Wellcome Sanger Institute, AI tools are dramatically accelerating some tasks within scientific pipelines, creating new bottlenecks where accelerated tasks feed into non-accelerated ones. This "jagged" acceleration is already straining peer review, funding decisions, and regulatory approval processes.

In the context of AI R&D, this means that while AI can speed up coding, experimentation, and analysis, other parts of the pipeline—such as human oversight, safety reviews, and coordination between institutions—may not keep pace. This could lead to either dangerous science proceeding unchecked or breakthroughs being held up, depending on how the bottlenecks are managed.

The Road Ahead: What This Means for AI Timelines

The METR report represents a significant update to our understanding of AI timelines. Many external researchers, including the AI 2027 team, have projected more moderate progress, but the evidence from Anthropic suggests that the frontier is moving faster than expected. If AI is indeed doubling the rate of AI progress, the implications for society are profound.

Cotra describes this report as "the most important update we have gotten about AI risk in the last couple of years." The possibility of superintelligence within a year, while uncertain, demands urgent attention from policymakers, researchers, and the public.

The industry must now grapple with the question of how to safely navigate this acceleration. The window for implementing robust safety measures may be closing rapidly. As Cotra warns, "We may be running out of time to get on the same page about timelines and takeoff speeds."

The METR report is a wake-up call. It provides concrete evidence that AI is not just a tool for other domains but is now actively accelerating its own development. The question is whether humanity can keep up—and whether it can do so safely.