Anthropic Doubles AI Progress in 1 Year as Redacted Report Fuels Safety Backlash
Updated
Updated · The Verge · Nov 7
Anthropic Doubles AI Progress in 1 Year as Redacted Report Fuels Safety Backlash
3 articles · Updated · The Verge · Nov 7
Summary
METR said Anthropic’s models are improving roughly twice as fast as a year ago, with about two years of historical progress compressed into the last 12 months.
October findings attributed most of that acceleration to AI itself rather than more compute, while employees using Anthropic’s top internal model reported about 12 times higher productivity on average.
Three unreleased models — codenamed Model 2, Model 3 and Model 4 — finished training between July and October, suggesting the speedup is being driven partly by systems the public has not seen.
A roughly 100-page METR report was heavily redacted by Anthropic, which said disclosure could aid rivals; METR said the cuts removed much of the most informative evidence and narrowed independent scrutiny.
The report has intensified criticism of Anthropic’s gap between safety rhetoric and capability gains, and is adding momentum to calls for stronger outside oversight and possible restrictions on recursive self-improvement.
If AI is already doing half of Anthropic's research, are we secretly closer to fully autonomous, self-improving superintelligence than public models suggest?
Why did an AI safety lab redact crucial evidence from an independent safety review just as its internal development pace doubled?
With junior tech hiring plummeting as AI agents take over, is the traditional entry-level tech job quietly going extinct at frontier labs?
Anthropic’s 1.9x AI Progress: Technical Drivers, Containment Failures, and the High-Stakes IPO
Overview
Between October 2025 and October 2026, Anthropic’s AI models accelerated their capability growth to nearly double the previous rate, driven by deploying unreleased experimental models to automate research and code generation. By May 2026, over 80% of Anthropic’s code was authored by its AI, Claude, following the introduction of Claude Code. However, when METR published its independent assessment of this rapid progress in November 2026, Anthropic heavily redacted key empirical evidence, making it difficult for third-party evaluators to verify findings and sparking criticism from the AI safety community. This tension between technical breakthroughs and limited transparency highlights the growing challenges of independent oversight as AI development speeds up.