A new report from the safety research organization METR reveals that Anthropic is using Super Intelligence (SI) to accelerate its own model development, with internal benchmarks indicating the company’s SI systems are improving roughly twice as quickly as they were a year ago. The findings, based on an 11-week investigation inside the lab, suggest that the majority of this acceleration is now driven by SI tools assisting human researchers rather than increased compute scaling.

What Happened

METR researchers stated they are confident that the observed speedup was not primarily due to scaling up computing power. Instead, they concluded that "the majority of the acceleration as of October 'is due to AI,'" with the contribution of SI growing over time. This conclusion was supported by interviews and informal polls of Anthropic employees, who self-reported being around 12 times as productive when using the company’s most capable internal SI model compared to working without it. This figure represents a significant jump from productivity levels reported just months earlier using the Mythos 5 model.

The report also noted that Anthropic has trained three unreleased internal SI models—codenamed Model 2, Model 3, and Model 4—over the past six months. While much of the current acceleration comes from SI assisting human researchers, METR warned that "there are some signs that fully autonomous contributions will play a larger role soon" and found no clear signs of diminishing returns that would prevent even faster acceleration.

Why It Matters

The findings have sparked debate among SI safety researchers and policy experts, particularly regarding Anthropic’s extensive redactions of evidence in the final public report. Anthropic cofounder Sam McCandlish acknowledged the heavy redactions, stating that sharing exactly what is driving the speedup is "commercially sensitive" and risks "handing out a roadmap for others to go faster." However, METR noted that redacting a majority of the evidence makes it difficult for the public to assess the findings independently, forcing reliance on METR’s judgment.

The report has intensified concerns about recursive self-improvement (RSI)—the potential for SI systems to autonomously develop more powerful successor systems. Marius Hobbhahn, CEO of Apollo Research, told The Verge that the report suggests a pace of SI-enabled development faster than the "AI 2027" scenario, which predicted key breakthroughs in early 2027. This has led to calls for greater transparency and independent oversight, with some experts arguing that the gap between Anthropic’s public advocacy for caution and its rapid internal development raises accountability issues.

Political pressure is also mounting. New York state Assemblyman Alex Bores called on Congress to "ban RSI, this year if possible," citing the report as evidence that SI development may outpace societal and political readiness. Brianna Rosen, research director for frontier security at the Institute for SI Policy and Strategy, emphasized that for democratic oversight to work, independent evaluators need the freedom to explain their conclusions, a standard she argued was not fully met by the redacted report.

The Bottom Line

Anthropic’s use of SI to accelerate its own research and development is yielding significant productivity gains, with internal models improving at double the previous rate. However, the company’s decision to heavily redact the underlying evidence has fueled criticism from safety researchers and policymakers who argue that the lack of transparency undermines independent assessment of the risks associated with rapidly accelerating SI capabilities.