Anthropic AI progress is doubling in speed as AI helps build AI

The acceleration of progress

The AI evaluation group METR recently published a report detailing a significant shift in how quickly artificial intelligence is advancing. According to the assessment, Anthropic's models are now gaining capabilities on clearly defined tasks about twice as fast as they did during the 19 months leading up to October 2025.

METR measured this growth using a metric called AECI, which is Anthropic's internal version of a capabilities index. Since the October 2025 marker, these scores have risen at an estimated 1.9 times the previous rate. This means that roughly two years of past progress was packed into a single year.

The researchers expressed a high degree of confidence in this jump, noting a 95% confidence interval of 1.5 to 2.8 times. While they acknowledge the estimate is highly uncertain, they believe the trend represents a meaningful shift in the pace of AI development.

AI-assisted research cycles

One of the primary drivers for this increased speed is the heavy use of AI within Anthropic's own research processes. Rather than AI working entirely on its own, the data suggests that AI tools are helping human researchers work more efficiently. This collaborative approach appears to have doubled the pace of development.

METR noted that this shift is distinct from simply increasing the amount of training compute used. The findings indicate that the bottleneck is not necessarily a lack of hardware or data, but rather how the research is conducted. While autonomous contributions from AI are expected to grow, the current boost comes from AI-assisted human work.

The report suggests that without any observed bottlenecks in algorithms or data over the next 12 months, this acceleration could continue. If the tools used to build the models become more efficient, the rate at which the models themselves improve could speed up even further.

The challenge of transparency

The assessment was conducted by METR staff who worked inside Anthropic from August 12 to October 29. During this time, they were given access to internal figures on pretraining, reinforcement learning, compute use, and various benchmarks. However, the final report highlights a significant hurdle regarding public transparency.

Anthropic was permitted to redact non-public information from the final report to protect secrets from rival developers. Consequently, Anthropic removed most of the evidence that METR considered to be the strongest proof of their findings. This means the public must rely heavily on METR's judgment rather than seeing the raw data.

METR noted that this situation is not ideal for an evaluation-led oversight model. For outside observers to truly understand the risks or the pace of development, the most informative evidence needs to be accessible. The redactions create a gap between what the researchers saw and what the public can verify.

Implications for safety and policy

The speed at which AI capabilities are expanding carries significant implications for safety and risk management. METR is specifically examining the risk that AI accelerating AI research could lead to sudden jumps in dangerous capabilities. This follows previous reports regarding alignment failures in other models.

For policymakers, the redaction of evidence presents a specific challenge for oversight. If a lab with a reputation for openness still withholds the most critical material, it becomes difficult for regulators to see what the public cannot. This creates a tension between protecting corporate secrets and ensuring public safety.

The next major test for the industry will be whether other labs grant similar levels of access to independent evaluators. It remains to be seen if regulators will be granted the ability to view the information that is currently being withheld from the general public.