A new analysis from AI safety nonprofit Model Evaluation and Threat Research (METR) suggests that Anthropic's AI models have been improving at a dramatically faster clip since October 2025—1.9 times faster than before. Even more striking, the group estimates that AI itself is responsible for roughly 80% of that acceleration.
The finding, first reported by Reuters, comes from an 11-week assessment that combined confidential research access, interviews, and performance data from unreleased systems. METR stresses that the 80% figure is uncertain, but says it is confident the speed-up wasn't simply the result of throwing more training compute at the problem.
If that conclusion holds, it hints at a significant shift: capability gains may be starting to decouple from raw infrastructure spending. Instead, progress could be coming from faster research cycles—better tooling, tighter experimentation loops, and AI assisting researchers in ways that compound over time.
What's driving the acceleration?
METR's analysis points to AI's role in speeding up the research process itself. Instead of researchers manually running experiments and analyzing results, AI systems can help generate hypotheses, write code, and evaluate outcomes, shortening the time between idea and test. That kind of iterative loop can compound quickly, leading to faster benchmark gains without a proportional increase in compute.
The group also said it didn't see convincing evidence of near-term bottlenecks that would naturally slow progress within the next year. However, it noted that today's systems still aren't making fully autonomous, high-impact contributions to the core stages of model training—meaning human oversight remains critical.
This dynamic creates a practical tension for frontier labs like Anthropic. When iteration speeds up, the "slow parts" of the pipeline become less about chips and more about everything around a release: safety testing, red-teaming (structured attempts to break a system), governance sign-offs, and monitoring after deployment. These steps can't be accelerated as easily as compute, potentially making them the new bottleneck.
What it means for investors
For markets, the METR finding could change the payoff from more compute. If benchmark progress can accelerate without a matching rise in training compute, then the usual assumption—that more GPUs equals proportionately better models—looks weaker at the margin. That matters for an industry planning hundreds of billions of dollars in data centers and chips. Returns may depend as much on research productivity as on hardware scale.
This is especially relevant given the massive capital commitments already in motion. For instance, Google has backed a $19 billion data center project tied to Anthropic, and the company's own IPO filing revealed an $895 billion compute bill—a staggering figure that underscores how much of the AI buildout is predicated on compute scaling.
If research productivity becomes a bigger driver of progress, the economics of that buildout could shift. Investors may start paying more attention to how efficiently labs use their compute, rather than just how much they buy.
The finding also shifts what limits frontier labs like Anthropic. Faster model cycles can pull revenue opportunities forward, but they can also make safety evaluations and release processes the choke point. That could affect launch timing, regulatory scrutiny, and how confidently big customers adopt new systems. For a company reportedly eyeing a valuation of $1.5–1.6 trillion, these operational constraints could be as important as model capability.
For everyday investors, the takeaway is that AI progress isn't just about hardware anymore. The ability to accelerate research cycles—and the safety and governance steps that come with it—may become key competitive advantages. That could influence which companies thrive in the next phase of the AI boom, and how much of the massive capital spending actually translates into better models and returns.
METR's findings are preliminary and based on a single lab, but they add to a growing debate about whether the AI industry's compute-heavy approach is sustainable. As more data emerges, investors will be watching closely to see if the trend holds—and what it means for the hundreds of billions of dollars already committed to the AI infrastructure buildout.
