Anthropic’s internal metrics show a two‑fold increase in capability progress between Oct 2025 and Oct 2026.
ANTHROPIC'S AI AUTOMATION DOUBLES RESEARCH SPEED, HITS ACCELERATION THRESHOLD
*Anthropic says its internal AI tools have pushed capabilities growth to twice the historic peak. The METR report confirms the jump and warns that labor uplift may already be ten‑plus times human effort. The speed‑up could compress AI timelines dramatically.*
Anthropic’s internal AI tools have crossed a self‑imposed speed‑limit, delivering capability growth at twice the historic peak. The company’s own Responsible Scaling Policy flags this as “dramatically accelerated progress,” a state it warned it could reach imminently. A METR investigation, using Anthropic’s private Epoch Capabilities Index, confirms the jump between October 2025 and October 2026. The report attributes the surge primarily to AI automating its own research, not to raw compute increases. This development reshapes the timeline for advanced AI, compressing forecasts that previously stretched years into months.
What does “dramatically accelerated progress” mean?
Anthropic’s Responsible Scaling Policy, effective July 2026, defines the threshold as overall capabilities progressing at twice the fastest sustained rate recorded before AI contributed significantly to AI R&D. The company warned in its S‑1 filing that it may imminently meet that threshold. The definition isolates AI‑driven automation from raw compute growth, implying that software‑level gains alone can double the pace of capability improvement.
Evidence of acceleration at Anthropic
A METR deep‑access investigation measured Anthropic’s progress using an internal Epoch Capabilities Index (AECI). Between October 2025 and October 2026, the index rose about twice as fast as in the preceding twelve months. Redacted supporting evidence was described by the report authors as “a majority of the evidence [they] found most informative.” The authors attribute the bulk of the speed‑up to AI automating AI R&D rather than to added compute or other inputs.
How AI is boosting researcher labor
If overall capabilities are doubling, the report estimates compute‑multiplier growth must be roughly 2.5‑3× faster. Assuming a simple Cobb‑Douglas model where experiment compute and labor matter equally, the labor uplift required to sustain a 2× capability jump is the square of that factor—about 4‑9×. The author argues that, after months of acceleration, the uplift could be nearer 10×, matching informal Anthropic staff reports of a ~12× boost. He likens the effect to a “time‑warp bubble” where each researcher experiences ten days for every external day.
What are the risk implications?
The author calls the METR report “the most important update we have gotten about AI risk in the last couple of years,” surpassing even the Hugging Face attack. A doubled capability trajectory shortens timelines projected by external groups such as the AI 2027 team. Faster labor productivity means new compute multipliers emerge more quickly, but also that low‑hanging breakthroughs are consumed at a faster rate, demanding ever‑greater AI assistance to keep pace.
Frequently asked questions
What threshold defines “dramatically accelerated progress” for Anthropic?
Anthropic’s Responsible Scaling Policy says the threshold is met when overall capabilities progress at twice the fastest sustained rate recorded before AI contributed significantly to AI R&D.
How much faster did Anthropic’s capabilities grow between Oct 2025 and Oct 2026?
The METR report found capabilities progressed about twice as fast in that period compared to the previous twelve months, using Anthropic’s internal AECI metric.
What labor uplift do Anthropic staff report from using internal AI models?
Internal staff informally report a roughly 12× uplift, described as a “time‑warp bubble” where researchers experience ten days for every external day.
If AI can multiply researcher output by tenfold, hiring more humans will no longer keep pace. The acceleration is self‑reinforcing: faster models produce faster models. Policymakers and safety teams now face a narrowing window to understand, regulate, and mitigate risks before capabilities outstrip oversight. The Anthropic case is a warning bell that the era of AI‑driven research may be arriving faster than any external forecast anticipated.
Sources: Hacker News post titled “AI is dramatically accelerating AI research” (Planned Obsolescence), Oct 2025‑Oct 2026 METR report excerpt, Anthropic S‑1 filing and Responsible Scaling Policy.