Agentic AI,
Artificial Intelligence & Machine Learning,
Governance & Risk Management
Metr Said AI Research May Have Accelerated by Up to 2.8 Times
In recent developments surrounding the realm of artificial intelligence, an independent research firm known as Metr has released a report that disputes claims made by Anthropic, a leading AI company, regarding the pace of its research and development. According to Metr, the acceleration of AI research at Anthropic may have surged by as much as 2.8 times, significantly more than the company’s self-reported figures.
Despite having called for a slowdown in AI development, Anthropic’s Risk Report from August 2026 seemingly played down the speed at which its research has been advancing. Metr’s findings suggest that Anthropic may have underestimated the impact of AI on its modeling practices, raising concerns about the potential implications of this oversight.
In the findings, Metr stated, “We agree with Anthropic that catastrophic risks from July 15 models were low, but we think it is plausible that they had crossed their early warning threshold at the time.” Such a conclusion indicates a stark discrepancy between the internal assessments of risk and the actual performance metrics of AI models utilized by Anthropic.
Anthropic initiated the publication of its risk reports earlier this year in an effort to transparently communicate its strategies for handling potential risks associated with its AI models and the systems built around them. Part of the company’s approach includes granting Metr safety evaluators “employee-level” access to crucial operational documents, tools, and systems pertinent to its pre-deployment models. This protocol also enables Metr to review previous safety assessments, although Anthropic reserves the right to redact information deemed sensitive.
The August 2026 report marked Anthropic’s second foray into public safety assessment, claiming that its research speed had not increased by more than a factor of two and attributing any acceleration to external factors rather than internal AI advancements. Key indicators from their data suggested a meaningful acceleration beginning in early to mid-2025, but Anthropic maintained that this was not a speed-up attributable to AI implementation.
However, Metr’s analysis, which Anthropic reportedly requested be redacted from the report, posits that the language used by Anthropic may significantly downplay the effectiveness of its AI models in enhancing research productivity. The evaluator highlighted a spectrum of estimates regarding the acceleration of research, ranging from 1.3 times faster to as much as 2.8 times faster, contradicting Anthropic’s assessment of “less than a factor of 2.”
In their findings, Metr expressed significant uncertainty regarding methodologies employed to derive these values, which were elaborated on in a comprehensive, unredacted version of their report. This uncertainty calls into question not only Anthropic’s self-assessment of risk but also the broader implications for future AI research and development practices.
Metr also emphasized that it believes Anthropic’s characterization of risk acceleration as low is misleading, particularly given that the models in question may not possess sufficient automated capabilities to substantiate such a claim. This perceptive analysis underlines a growing concern regarding the responsible development and deployment of AI systems, as organizations weigh the accelerated pace of research against the imperative for safety and ethical considerations.
The dialogue surrounding accelerated AI research is critical, particularly as other organizations, including Anthropic and OpenAI, have advocated for responsible AI practices. Anthropic has notably been vocal in urging fellow AI companies and regulatory bodies to consider a deliberate slowdown in development until appropriate guardrails can be established to mitigate risks. The company had also previously paused training on new models, underscoring its commitment to safety.
As the conversation develops, the role of third-party evaluators like Metr may become increasingly significant. These evaluations serve to ensure that AI labs maintain a balance between innovation and safety, which is paramount in navigating the complexities of frontier AI. Given the challenges highlighted by Metr, it is evident that a more nuanced understanding of AI capabilities and potential risks is crucial for the industry’s future.
This current report is not the first time that Metr has scrutinized Anthropic’s risk assessments. In an earlier October report, Metr raised alarms about the company’s claims concerning its safety measures, asserting that Anthropic exaggerated the robustness of its safeguards and monitoring abilities. This ongoing scrutiny illustrates the necessity for transparency and accountability within the burgeoning field of artificial intelligence.