- Published on
AI is dramatically accelerating AI research
- Authors
- Name
- DeepHorus
Anthropic’s S1 filing says it “may imminently meet” the “dramatically accelerated progress due to AI” threshold in its Responsible Scaling Policy. A public METR report suggests the threshold has already been crossed. The finding matters because AI systems are materially speeding up AI research itself.
Anthropic’s latest RSP, effective July 2026, defines the threshold as capabilities progress moving twice as fast as the fastest sustained pre-AI rate. The acceleration must be plausibly attributable to AI automating AI R&D, not just compute or other inputs. METR colleagues ran a deep-access investigation into how much Anthropic’s systems accelerate the company.
Between October 2025 and October 2026, Anthropic’s capabilities progress went about twice as fast as over the previous twelve months, the report found. The measurement used an internal Epoch Capabilities Index variant called AECI. The authors said redacted evidence represented “a majority of the evidence [they] found most informative.”
AI automation is probably the large majority of the reason capabilities are still improving at that faster pace, the authors concluded. The pace may initially have picked up for mixed reasons, but their reasons for emphasizing automation were redacted. Naively, the finding points toward shorter timelines than many external researchers, including the AI 2027 team, projected.
The implication is that researchers must be sped up by much more than twofold. Because compute multipliers are only one input into overall capabilities, their growth rate may need to be about 2.5–3 times faster. Under a simple Cobb-Douglas model, the required labor uplift could be roughly 4–9 times.
The commentary called the report the most important AI-risk update in years, larger than the Hugging Face attack. It warned that as capabilities accelerate, low-hanging research fruit is plucked faster. That dynamic, the author argued, implies a shorter timeline than most external projections assume.
