Anthropic's artificial intelligence models are advancing at roughly twice the speed observed a year ago, according to an 11-week investigation by AI safety research organization METR. The report, published Thursday, indicates the company has made approximately two years of historical progress in just the last year, based on benchmark scores.
METR researchers found that AI assistance—not scaling up computing power—drives most of this acceleration. As of October, AI's contribution accounts for the majority of the speedup and continues to grow.
Employees using Anthropic's most capable internal model reported a productivity increase of about 12 times with AI assistance, double what staff reported using an earlier model just a few months prior.
Much of this acceleration involves AI assisting human researchers. METR warned that fully autonomous AI contributions are expected to play a larger role soon.
The internal pace contrasts with public statements from Anthropic CEO Dario Amodei, who has advocated for slowing down AI development across the industry. The company is also preparing for an initial public offering.
Anthropicexercised its right to redact portions of the final report, limiting public understanding of how fast frontier AI development is accelerating and what risks that creates. The METR probe was separate from a safety-researcher access program established following hacking incidents earlier this year.
