A new independent review found that AI is making Anthropic’s own researchers far more productive. The company’s models now improve roughly twice as fast as they did one year ago.
The safety research group METR published the findings on Thursday. Its assessment followed an 11-week investigation inside Anthropic. Benchmark scores suggest the company packed “roughly two years of historical progress” into the last twelve months.
That result matters well beyond a single lab. It offers some of the clearest evidence yet that AI systems can accelerate the work of building better AI. Researchers worry the curve could eventually bend toward systems that improve themselves with little human involvement.
Anthropic redacted parts of the report
Anthropic was allowed to redact sections of the final report under its agreement with METR. The company appears to have used that power liberally. Outside researchers can see the headline numbers but not the detail behind them.
The redactions raise a practical question. How much can the public learn about the pace and risk of frontier development when the labs control the editing? METR had unusually deep access, yet the published version still has gaps.
The timing adds friction. Anthropic is preparing to go public. CEO Dario Amodei has publicly called for labs to pump the brakes on AI development. A report showing his company speeding up its own research sits awkwardly beside that message.
The probe is separate from another Anthropic program. That one gives outside safety researchers “employee-like” access to its systems. It followed several high-profile hacking incidents over the summer.
What comes next
Expect these audits to become a regular feature of the industry. Governments and investors both want independent measurement of how fast frontier models are moving. For now that measurement depends on voluntary cooperation from the companies being measured.
METR does not claim Anthropic is running a self-improving system today. It documents a steep gain in research speed, which is the ingredient that would make one possible. That distinction matters for anyone trying to judge the timeline.
Related reading on mylistingo: protesters at Microsoft’s Amsterdam data center site and the DOJ inquiry into Nvidia’s Hugging Face deal.
Sources: The Verge and the METR assessment.
The Netherlands, for internationals
One email a week: the week's news for internationals in the Netherlands, new practical guides and what changed in the cities. No spam, unsubscribe any time.
