On Sept. 12, an essay published by Anthropic Chief Executive Officer Dario Amodei urged the global artificial intelligence sector to slow the pace at which the industry improves model capabilities. Yet inside the San Francisco-based company, which is preparing to go public later this month in what could be the biggest initial public offering of stock ever, development momentum has moved in the opposite direction.
An audit by Berkeley, California–based nonprofit Model Evaluation and Threat Research revealed that Anthropic is developing its artificial intelligence systems around twice as quickly as it was a year ago. Public appeals for a regulatory and industry-wide pause mask an intense operational acceleration ahead of the company's Wall Street debut, raising questions about whether internal development cycles are deliberately outpacing external governance.
Internal Velocity Versus Public Caution
METR conducted its evaluation after Anthropic invited the research group to inspect internal data and interview personnel over an 11-week window running from mid-August through late October. The assessment specifically examined whether the growth rate of capabilities was compounding and how much of that gain stemmed from automated workflows.
In a Sept. 12 essay responding to mounting concerns about runaway AI, Anthropic CEO Dario Amodei urged the AI world to "slow the pace at which we improve the capabilities of AI models."
Yet during the very period the independent assessors spent reviewing internal records, Amodei called for the industry to slow down so safety measures could catch up. METR discovered that Anthropic's own public risk report published in August understated the actual speedup when contrasted with private documentation and nonpublic data supplied to evaluators.
The Feedback Loop of Self-Improving Systems
The mechanics behind this metric shift are structural rather than tied to simple hardware procurement. METR was able to confidently rule out the expansion of computing power used in training as the primary driver of the acceleration. Evaluators determined that physical bottlenecks, including compute restrictions or data limits, were not holding back Anthropic's AI progress. Instead, Anthropic confirmed that its Claude chatbot increasingly assists the enterprise in engineering newer, more advanced iterations of itself.
Evaluating the findings on X, Anthropic co-founder and responsible scaling officer Sam McCandlish agreed with many of METR's conclusions, observing that artificial intelligence is speeding up internal research even if the precise scale remains difficult to isolate. McCandlish noted that the rate of progress introduces significant safety considerations because it determines how rapidly risks expand. At the same time, Anthropic applied redactions to the findings shared with the public.
Sam McCandlish acknowledged that redacting the records makes it harder for outside parties to check METR's work, but defended the choice by explaining that revealing "exactly what is driving the speedup isn't the right thing to do" because it is "commercially sensitive" and "risks handing out a road map for others to go faster."
This commercial defense addresses the fundamental objection raised by assessors. METR conceded that its headline estimate could overstate systemic acceleration because the measured double-speed performance might be limited strictly to well-specified tasks. By keeping the technical details behind the speedup redacted, Anthropic shields its engineering advantages from competitors while preventing outside auditors from verifying whether the feedback loops extend beyond narrow operational benchmarks.
Independent evaluation frameworks face structural limits when corporate targets take precedence. Amodei urged the market to hit the brakes on capability gains while Anthropic's own engineering teams doubled their speed to meet Wall Street expectations.
What it means
Heavy clusters of fresh servers had nothing to do with it — the company simply tasked Claude with building its own successors while METR auditors tried to reconcile redacted files with public warnings. With a projected valuation of $965 billion on the line, preachments about pausing look less like civic caution and more like a clever way to stall rivals before the opening bell. Can Amodei persuade public markets that safety protocols remain intact while Anthropic censors the very audit evaluating its accelerating models?
Our stories: Anthropic $965 Billion IPO: High Valuation vs Chinese Rivalry (August 12); Anthropic S-1: $8B Loss and $518B Compute Bill Before $2T IPO (September 29).