Independent Assessment of AI R&D Acceleration at Anthropic in October 2026

metr.org49 min read

Between August and October 2026, METR staff conducted a high-access evaluation at Anthropic to assess whether AI is accelerating AI R&D. METR researchers concluded that Anthropic is likely improving model capabilities on well-specified tasks at roughly twice the historical rate, with a majority of this acceleration attributed to AI usage, though Anthropic redacted key evidence and significant uncertainties remain. Open article →

Notable comments

HN discussion focused on skepticism about redaction-heavy evidence, uncertainty in the reported ~2x AI-driven R&D acceleration, and side critiques about real-world frontier model regressions and broader existential concerns that received little attention.

HN · m_w_ Critiques the article’s framing as opening with “Anthropic’s redacted the best parts” and notes the disclosed/partly-disclosed evidence is “dominated by no/partly” sources, undermining trust.

HN · jackschultz Suggests the cited “acceleration in late 2025” may be largely explained by a non-AI threshold effect tied to Opus 4.5, where tests stopped working because the model became so strong that ignoring it outperformed prompting.

HN · kian Provides first-hand workflow evidence that a newer model (Opus 5.5) has caused significant operational breakage—ignoring instructions, inventing requirements, silently breaking code, and refusing fixes—suggesting capability/behavior isn’t monotonically improving.

More on this topic