Physix Frontier badgePhysix Frontier
PublishChinese | EN

Physix Frontier · News Briefing Card (Hacker News · Nov 6, 2026)

METR: Anthropic AI R&D sped up ~2x, but key evidence redacted

KEY FACTS

  • METR conducted a high-privilege evaluation of Anthropic from August to October 2026, reviewing data on its AI R&D process.
  • METR estimates Anthropic's model capability improvement rate is about 2x the historical rate, with a range of 1.3x to 3x.
  • Since October 2025, the growth rate of Anthropic model benchmark scores has been 1.9x the prior rate, with a confidence interval of 1.5x to 2.8x.
  • METR attributes the acceleration mainly to Anthropic's extensive use of AI-assisted R&D, not to expanded training compute.
  • Anthropic redacted most of the key evidence; METR says the public cannot independently judge it and must rely more on Anthropic's conclusions.

KEY DATA

~2x (range 1.3-3x)Capability improvement rate multiplier
1.9x (95% CI: 1.5-2.8x)Benchmark score growth multiplier
August 12 to October 29, 2026Evaluation period

PHYSIX OBSERVATION

There is preliminary evidence of AI R&D self-acceleration, but key data is obscured by commercial confidentiality, leaving public oversight empty. If the acceleration is real, the safety window narrows; the industry needs a verifiable third-party audit mechanism, otherwise risk judgments can only rely on vendors' own accounts.

Source: Hacker News report