Anthropic AI development speed has doubled driven by AI assistance, METR report says

What happened

A new report published Thursday by the nonprofit AI safety research organization Model Evaluation and Threat Research (METR) reveals that the rate at which Anthropic's AI models are improving has roughly doubled over the last year (theverge.com, techtimes.com). According to The Verge, the 11-week investigation involved deep access to Anthropic's internal research data, benchmark scores from unreleased models, and employee interviews (theverge.com, the-decoder.com).

METR estimated that capabilities progress on well-specified tasks jumped from about 15 points per year before October 2025 to roughly 29 points per year afterward (techtimes.com, ijr.com). The nonprofit's evaluators stated they are confident this acceleration was not simply caused by scaling up the compute power used to train the models (theverge.com). Instead, METR estimates that approximately 80 percent of the speedup stems from AI being used to assist in the company's own research (the-decoder.com, ijr.com).

Why it matters

The findings offer clear evidence of artificial intelligence accelerating its own development, a key early warning signal in the safety debate (the-decoder.com). Industry observers warn that such recursive self-improvement could create a feedback loop leading to the rapid, sudden emergence of extreme capabilities (the-decoder.com, fastcompany.com).

Anthropic's own Responsible Scaling Policy sets a "2x Acceleration Threshold" for dramatic acceleration that should trigger warning levels if progress roughly doubles due to AI automation (the-decoder.com, techtimes.com). While Anthropic's public risk report stated the acceleration was less than a factor of 2, METR concluded that argument was "overstated" and that crossing the threshold by mid-July was plausible (the-decoder.com, techtimes.com).

Context

The evaluation arrives as Anthropic prepares to go public and as its CEO, Dario Amodei, has publicly called for AI labs to slow down development so safety measures can keep pace (theverge.com, fastcompany.com). Amodei noted in a September essay that recursive self-improvement "could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all" (fastcompany.com).

METR noted that Anthropic holds unilateral redaction rights over all non-public information and heavily censored the public version of the roughly 100-page report (the-decoder.com, techtimes.com). Anthropic described the redactions as "aimed at removing confidential information that could accelerate other AI developers" (the-decoder.com, techtimes.com). Because of these heavy omissions, the public must rely largely on METR's judgment rather than the underlying data (theverge.com, techtimes.com).

What we don't know yet

The report highlighted significant uncertainties, with METR noting its robustness checks span a wide range from 1.1x to 2.9x depending on modeling choices (techtimes.com). Furthermore, METR canceled a planned component of the audit that was meant to verify how Anthropic divides its compute between capabilities and safety work, after Anthropic removed that scope to focus on risk assessment (the-decoder.com, techtimes.com).