Anthropic announced this week in its S1 filing that it may have crossed a critical threshold defined in its Responsible Scaling Policy: “dramatically accelerated progress due to AI.” According to a recent METR report on AI research and development acceleration at Anthropic, this milestone appears to have already been reached.
The report found that between October 2025 and October 2026, capabilities progress at Anthropic advanced approximately twice as fast as the previous twelve months, measured using an internal version of the Epoch Capabilities Index. Researchers attributed this acceleration primarily to AI automating AI research and development tasks, rather than to increases in computing power or other inputs.
The implications extend beyond simple doubling. Because compute multipliers—algorithmic and architectural innovations—represent only one input to overall AI capabilities, a twofold acceleration in capabilities would require compute multipliers to advance roughly 2.5 to 3 times faster. Researcher labor is similarly just one input to discovering these multipliers; accounting for the need to test ideas through experiments, a simple economic model suggests labor productivity would need to increase roughly 4 to 9 times to produce a twofold acceleration in overall capabilities.
As capabilities accelerate, the rate of discovery becomes more demanding—“ideas are getting harder to find.” To maintain accelerated progress without proportionally scaling compute, labor growth rates would need to accelerate beyond historical rates. Since human workforce growth remains constrained, the productivity boost from AI would need to expand over time. Current estimates suggest AI is providing roughly a 10-fold uplift in researcher productivity at Anthropic, with informal reports from staff describing a 12-fold uplift from internal models.
This acceleration could compound rapidly. As AI capabilities improve further, researchers may experience increasingly dramatic speedups—progressing from month-long tasks completing in weeks, then days. If this feedback loop sustains without encountering steep diminishing returns or bottlenecks, progress could accelerate from 2x to 4x to 8x over successive periods. The METR report authors found limited evidence of bottlenecks that would halt this compounding effect, though they noted substantial uncertainty and redacted evidence suggesting acceleration may already exceed 2x and be increasing.
Key facts
- Anthropic’s capabilities have progressed roughly twice as fast over October 2025 to October 2026 compared to the previous year
- The acceleration is primarily attributed to AI automating AI research tasks rather than increased computing resources
- METR researchers estimate AI is providing approximately a 10-fold productivity uplift to human researchers at Anthropic
- The report identified limited evidence of bottlenecks that would prevent further compounding acceleration