Anthropic co-founder Sam McCandlish has defended the company’s decision to redact much of an independent assessment of its AI research, arguing that disclosing the specific factors behind its progress could reveal commercially sensitive information and give rival labs a roadmap for moving faster. He made the case in a post on X after METR published its findings, as reported by The Verge’s November 7 coverage.
METR’s 11-week assessment concluded that Anthropic was likely improving model capabilities on well-specified tasks at around twice the rate seen between March 2024 and October 2025. Its benchmark analysis estimated growth at 1.9 times the earlier rate since October 2025, with a 95% confidence interval of 1.5 to 2.8 times. The researchers said most of the current acceleration was likely attributable to AI use in research and development, while stressing that the estimate carries substantial uncertainty. The group’s public assessment says Anthropic redacted a majority of the evidence METR found most informative, making it harder for readers to assess the conclusions. Abijita covered the original findings in its report on METR’s AI R&D assessment.
McCandlish acknowledged that the redactions made the work harder for others to check, but said Anthropic believed releasing the details would create commercial risks and could worsen the very acceleration concerns raised by METR. He also said the company plans to publish its own analysis of AI progress and additional safety measures in the coming weeks. Cybernews separately reported McCandlish’s explanation and METR’s criticism of the omissions.
The Verge reported that outside AI safety researchers questioned whether the public report provides enough evidence for independent scrutiny and called for deeper access to frontier AI labs. Anthropic declined to answer the outlet’s questions about the assessment and directed it to McCandlish’s post.
If this article helped you, please consider supporting our work. Every small contribution keeps Abijita.com independent and running.
