AI Is Already Doubling Anthropic's R&D Speed
Original: AI is dramatically accelerating AI research
Why This Matters
A verified 2x capability acceleration at a frontier lab resets assumptions about AI development timelines.
METR's deep-access report found Anthropic's AI capabilities progress doubled in speed between Oct 2025–Oct 2026 vs. the prior year, with AI automation cited as the primary driver — triggering Anthropic's own RSP acceleration threshold.
Anthropic disclosed in its S1 filing that it 'may imminently meet' the threshold for 'dramatically accelerated progress due to AI' as defined in its Responsible Scaling Policy. That threshold is crossed when capabilities progress runs at twice the fastest sustained pre-AI-contribution rate, attributable to AI automating AI R&D rather than raw compute gains.
METR, the AI safety evaluation organization, conducted a months-long deep-access investigation at Anthropic. Its report — parts of which are redacted, including what the authors call 'a majority of the most informative evidence' — found that the threshold has effectively already been crossed. Using an internal metric called AECI (Anthropic's version of Epoch's Capabilities Index), researchers found progress roughly doubled in pace year-over-year.
Author Ajeya Cotra (writing independently, not for METR) draws out the implied math: if overall capabilities doubled due to software progress, compute multipliers — the algorithmic improvements researchers produce — must have accelerated by 2.5–3x. Under a Cobb-Douglas model treating compute and researcher labor as roughly equal inputs, the implied speedup in effective researcher labor could be 4–9x. Cotra calls this report 'the most important update on AI risk in the last couple of years,' noting it implies timelines significantly shorter than most external forecasters, including the AI 2027 project, have projected.