An independent assessment of Anthropic’s internal materials and employee interviews says the company’s AI research may be approaching—and could have reached in July—its internal warning threshold for accelerated development. The assessment says the risk of catastrophic harm from faster AI progress remains low, but questions Anthropic’s case that the threshold had not been crossed, citing disputed assumptions and information that has not been fully disclosed. The findings come as the company prepares for an IPO, bringing renewed attention to its safety oversight and pace of development.
The assessment was published in early November by Model Evaluation and Threat Research (METR), a California-based nonprofit research organization. In an August safety assessment, Anthropic said the risk its models had crossed the threshold was low as of July 15. METR revisited the company’s public report alongside some internal risk materials Anthropic provided.
METR puts capability growth at about 1.9 times the historical rate
Anthropic’s internal warning threshold is a twofold increase in models’ ability to complete certain research and development tasks with less human involvement. The benchmark is intended to represent AI progress at a pace equivalent to compressing the advances made from 2018 to 2024 into a single year.
Based on changes from March 2024 to October 2025, METR estimates that Anthropic’s model capabilities have been improving at about 1.9 times the historical rate. The nonprofit says an estimate implied by Anthropic’s unpublished risk report is 1.97 times. Other reasonable measurement choices produce estimates ranging from 1.3 to 2.8 times, underscoring how much the result depends on methodology.
METR argues that Anthropic’s public case that it had not passed the 2x threshold was stronger than the evidence warranted. It says the threshold may have been reached by July 15. But the organization also cautions that Anthropic redacted some information and data, limiting the scope of outside analysis. Choices such as how clearly tasks are defined can also affect estimates of capability growth.
AI tools are contributing to research gains
METR estimates that about 80% of Anthropic’s growth in October came from AI systems helping conduct AI research, rather than from adding or expanding computing capacity. The finding points to researchers’ use of AI tools as one contributor to faster development, alongside investment in compute.
Anthropic employees have said that using the company’s internal models made them 12 times more productive. METR also says the company’s quarterly code output rose to about eight times its 2021–2025 benchmark. Those figures illustrate the effect internal tools may have on research efficiency, though the relationship between productivity and overall model capability depends on how each is measured.
METR says it has not identified clear bottlenecks or marked diminishing returns, raising the possibility that capability gains could accelerate further over the next 12 months. It also acknowledges that its limited perspective may have left constraints on research speed unaccounted for. The assessment is not a definitive forecast.
IPO brings safety commitments and development speed into focus
Anthropic CEO Dario Amodei has repeatedly called for a slower pace of technological development. In September, he warned that AI could pose serious risks and said capability gains might outpace improvements in safety measures. OpenAI CEO Sam Altman and other industry figures have also discussed the possibility of future AI systems developing autonomously beyond human control.
METR says higher research efficiency could help ease concerns across the AI industry about growing demand for computing resources and rising operating and development costs. At the same time, faster capability growth renews questions about whether safety measures can keep pace. METR has worked with Anthropic, OpenAI and other AI labs on assessments and research, and has taken part in independent testing of AI systems.
Anthropic is scheduled to set its IPO price on Wednesday and begin trading on Thursday. The company is reportedly seeking a valuation of up to $1.4 trillion and plans to list on the Nasdaq Global Select Market under the ticker ANT. With gaps in the available data and disagreements over methodology, the difference between Anthropic’s public assessment and METR’s analysis remains a point of scrutiny for investors examining the company’s safety governance ahead of the listing.