Anthropic models improve twice as fast as AI learns to build itself

AI is now building AI. An independent audit reveals Anthropic models are improving twice as fast, driven by AI doing the research itself.

Models · Source: Reuters

What happened

AI safety nonprofit METR just dropped a bombshell report after an 11-week internal audit of Anthropic. The assessment ran from August to October and drew on confidential research and unreleased models. The finding is stark. Anthropic benchmark capabilities are improving roughly twice as fast as they did between March 2024 and October 2025. Specifically, the growth rate hit a 1.9 multiplier. This acceleration is happening right as the San Francisco company prepares for a massive IPO later this month.

The speedup is not coming from throwing more computing power at the problem. METR confidently ruled out compute scaling as the cause. Instead, researchers point to recursive self-improvement. Anthropic is using its own Claude chatbot to write code, run experiments, and develop its next generation of models. Informal employee polls suggested productivity improvements of up to 12 times. While AI cannot yet make fully autonomous contributions to the most important training stages, Anthropic researchers noted significant signs of life implying that milestone is approaching fast.

We cannot see the full picture. Anthropic redacted key details from the report, limiting how METR could communicate its estimates. The nonprofit noted high uncertainty, suggesting the true acceleration could be anywhere from 1.5 to 2.5 times the historical rate. METR also stated that Anthropic public risk reports understate the actual level of acceleration. Ironically, this news breaks just weeks after Anthropic CEO Dario Amodei publicly called for the industry to slow down advanced AI development to let safety measures catch up.

Key facts

Why it matters

The timeline for AI capabilities just compressed violently. If models improve at double the historical rate without needing proportional increases in compute, the barrier to creating highly capable agents drops. The industry has spent hundreds of billions of dollars building computing infrastructure under the assumption that compute equals progress. If AI itself becomes the primary driver of technological progress, the economics of model training shift entirely. Builders relying on current model limitations to defend their product moats will find those moats evaporating much faster than anticipated. You can no longer plan product roadmaps around a linear progression of AI capability.

This marks the beginning of true recursive self-improvement in commercial AI. If AI can autonomously accelerate its own research and development, the bottleneck shifts from human engineering hours to raw data and oversight. This dynamic will force regulators and safety organizations into a corner. They are currently trying to police human-driven development cycles. They lack the frameworks to monitor AI systems that autonomously iterate on their own codebases. The fact that an independent evaluator found no bottlenecks holding back Anthropic progress means the runaway AI scenario is no longer just a theoretical risk.

For builders

Stop building for current model limits

Models are advancing at nearly double the historical rate. If your startup solves a problem that Claude currently struggles with, Claude will likely solve it natively in the next release. Build products that assume near-perfect reasoning capabilities are just months away.

Leverage AI for your own R&D

Anthropic employees reported massive productivity gains by using AI to write code and run experiments. If the top AI labs are using AI to build faster, you must do the same. Founders who refuse to integrate AI into their core engineering loops will lose to teams that do.

Prepare for regulatory whiplash

Anthropic leadership is calling for a slowdown while the company accelerates toward an IPO. This contradiction invites heavy government intervention. Startups building autonomous agents should expect sudden compliance costs as regulators panic over recursive self-improvement.

My take

We think humans are building AI, but AI is already the builder. Anthropic hid the strongest proof behind redactions, but the math is clear. The ground is moving twice as fast. If you are just learning to use AI tools instead of learning to build alongside them, you are already obsolete.

Original reporting: Reuters. This is my rewrite and opinion.

More AI news for builders