Anthropic is developing its artificial intelligence systems at approximately twice the speed it managed a year ago, according to a new analysis by the independent nonprofit Model Evaluation and Threat Research (METR). This acceleration occurs despite CEO Dario Amodei’s recent public plea in a September 12 essay for the industry to "slow the pace at which we improve the capabilities of AI models" to allow safety measures to catch up. The findings, released Thursday, provide rare external visibility into internal data during a period when the Trump administration has encouraged AI companies to self-police on safety matters.

The Recursive Self-Improvement Loop

METR researchers attribute the doubling of development speed primarily to AI assisting in its own creation. Anthropic has acknowledged that its Claude chatbot is increasingly helping the company develop newer, more advanced versions of itself. The nonprofit’s assessment, based on 11 weeks of internal data from mid-August through late October and employee interviews, confidently ruled out increased computing power as the driver. Instead, the acceleration appears to stem from efficiency gains and recursive improvements in the training process itself.

Redactions and Commercial Sensitivity

The transparency of these findings is limited by Anthropic’s decision to redact key details. Sam McCandlish, Anthropic’s co-founder and responsible scaling officer, responded to the report on X, stating that while many of METR’s conclusions align with his own, sharing the exact drivers of the speedup was not "the right thing to do." He argued that such details are "commercially sensitive" and risk "handing out a road map for others to go faster." Consequently, METR noted that the public risk report published in August "overall understates the level of acceleration" actually occurring within the lab.

Implications for the Upcoming IPO

This analysis arrives as Anthropic prepares for a Wall Street debut later this month, potentially the largest initial public offering in history. The report highlights that there are currently no "bottlenecks" such as compute or data restrictions holding back progress. While METR cautioned that AI is not yet making fully autonomous contributions to the most critical parts of model training, researchers found significant "signs of life" suggesting this autonomy is imminent. Amodei himself has defined this trend as "recursive self-improvement," a concept that raises urgent questions about control as the company scales toward public market scrutiny.

Key Takeaways

  • Anthropic’s development speed has doubled year-over-year, driven by AI-assisted research rather than compute increases.
  • CEO Dario Amodei’s call for a slowdown contrasts sharply with the rapid internal acceleration documented by METR.
  • Anthropic redacted specific data on what drives the speedup, citing commercial sensitivity and competitive advantage.
  • The report confirms no current bottlenecks in compute or data, with AI autonomy in training processes appearing imminent.

The Bottom Line

Anthropic is sprinting toward an IPO while preaching caution, and their redacted data suggests the real acceleration might be even more aggressive than the public narrative admits.