Anthropic has reached a threshold where capabilities progress is going about twice as fast as it had been previously, potentially due to AI automating AI R&D. According to a report from planned-obsolescence.org, capabilities progress at the company between October 2025 and October 2026 moved approximately twice as fast as the preceding twelve-month period.
Anthropic RSP Thresholds Signal Imminent Acceleration
The acceleration follows the internal metrics established in Anthropic’s Responsible Scaling Policy (RSP). Under the July 2026 version of the policy, a “dramatically accelerated progress due to AI” threshold is triggered when capabilities progress occurs at twice the fastest sustained rate observed before AI contributed significantly to R&D. The investigation into Anthropic’s internal progress indicates this milestone has already been crossed.
This development marks a departure from the May 2025 version of the RSP, which defined the AI R&D-5 risk threshold through a somewhat weaker milestone. The current data suggests that the transition from human-led discovery to AI-augmented R&D is moving faster than previously projected by external research teams.
Anthropic updated its S1 this morning. Notable additions re automated AI R&D and human oversight of it: "[…] our internal estimates and our expectations regarding future models, rather than our current models, indicate that we may imminently meet our Responsible Scaling… pic.twitter.com/i7tImT2iz6
The Mathematical Mechanics of Research Uplift
The doubling of overall capabilities does not imply a simple 2x increase in human labor efficiency. Instead, the relationship between compute, software innovation, and human effort creates a compounding effect. In the context of AI development, researchers focus on discovering “compute multipliers”—software innovations and architectural efficiencies that allow more capability to be extracted from the same amount of training compute.
If overall capabilities are advancing twice as fast due to software progress, the rate of growth in these compute multipliers must be significantly higher than 2x. Based on the mechanics of R&D, the required uplift to human labor to sustain this new pace is estimated to be substantially higher. While a simple model might suggest a 4x to 9x speedup in labor, current informal reports from Anthropic staff suggest a 12x uplift from the use of internal models. This does not mean researchers are working ten times more hours; it means they are operating in a “time warp” where AI allows them to process, read, and implement ideas at a vastly accelerated frequency.
This report on AI R&D automation (RSI) is a big update towards ASI within 1-2 years. It says AI progress is sped up by RSI right now by around 2x. In AI 2027, this happens around March 2027. I was seriously wrong about RSI speed up; I previously expected much less in 2026… https://t.co/1qKnCrrENK
Estimating the Acceleration Multipliers
| Metric Component | Observed/Estimated Impact |
|---|---|
| Overall Capabilities Progress | ~2x acceleration |
| Required Compute Multiplier Growth | Estimated 2x to 3x faster |
| Required Human Labor Uplift | Estimated 4x to 12x faster |
AI Progress Could Accelerate Through Compounding Returns
If increasing labor uplift is strong enough to overcome diminishing returns, the rate of AI progress could start going 4x faster, and then 8x, and then 16x.
The report notes that investigators found no immediate evidence of bottlenecks or steep diminishing returns that would halt this compounding acceleration. Instead, redacted evidence indicated that acceleration might already be exceeding 2x and increasing over time. This trajectory suggests that the current rate of progress could lead to a state where AI systems possess the capability to develop novel weapons or overpower human-controlled systems if misalignment occurs.
Security Deficits in the Face of Rapid Takeoff
Under the current RSP, reaching the AI R&D-5 threshold requires security measures capable of protecting against model-weight theft by state-level adversaries. However, the industry’s current security posture remains a point of vulnerability.
- Model Weight Security: Recent breaches at both OpenAI and Anthropic by amateur groups highlight that current defenses against weight theft are insufficient for the coming scale of risk.
- Algorithmic Secrecy: Protecting the specific architectures and training methodologies that drive these multipliers is as critical as protecting the weights themselves.
- Governance Infrastructure: There is a critical need for robust logging, universal monitoring, and countermeasures against monitor collusion to manage the risks of an intelligence explosion.
The transparency of the data remains a significant hurdle for the broader scientific community. Because much of the most informative evidence in the METR report was redacted, external researchers cannot independently verify the exact speed of the acceleration or the specific mechanisms driving it. This lack of transparency makes it difficult to align global safety standards with the actual velocity of the technology.
