Watchdog Report Highlights Faster AI Progress at Anthropic

A recent watchdog report released by independent evaluation group METR indicates that Anthropic's AI models were likely improving on specific tasks roughly twice as fast in late 2026 compared to earlier periods. The November assessment utilized AECI measurements, which tracked capability growth rates through snapshots of unreleased models.

METR interpreted this trend as representing about two years of historical benchmark progress achieved within a single year. While the group judged that AI assistance likely accounts for most of this acceleration, Anthropic redacted much of the underlying evidence. This decision left outsiders with a clear view of faster benchmarks but less visibility into the exact causes behind them.

Understanding the Watchdog Report Findings

The evaluation noted that non-AI factors also played a significant role in late 2025, though AI assistance now drives the majority of the progress. Furthermore, METR ruled out an acceleration in training compute growth as the primary explanation, although detailed supporting data for that exclusion was withheld due to commercial sensitivity.

Anthropic co-founder Sam McCandlish acknowledged that the redactions made independent checks more difficult, while defending the decision to protect proprietary details. Meanwhile, METR's companion review of Anthropic's August Risk Report questioned whether an early-warning acceleration threshold had officially been crossed, noting that direct catastrophic risks from tested models remained low.

AI models are improving at twice their historical pace, according to a recent independent evaluation group assessment.

Why it matters

The findings highlight rapid advancements in artificial intelligence capabilities and raise questions about how quickly progress could compound. Understanding these developments is crucial for tracking the future trajectory of machine learning technology and its broader implications.