Anthropic has often pushed for careful AI rules, testing and global coordination before powerful systems become too risky. But a new independent review says the company’s own AI progress is now moving roughly twice as fast as before. The finding creates an uncomfortable gap between its public safety message and its private research pace.
METR, an AI-evaluation nonprofit, spent 11 weeks studying Anthropic’s internal data and interviewing staff. It estimates that Anthropic is improving its models on well-defined tasks at about double its earlier rate, though the true figure could range from 1.3 times to 3 times.
The striking part is why: around 80% of the speed-up appears to come from Anthropic using its own AI to help build better AI. METR also found no clear bottleneck slowing this cycle, while the company kept much of the strongest evidence secret to avoid giving rivals a roadmap.
This does not mean Anthropic has abandoned safety. But it shows that “pacing the frontier” is easier to promise than to practise when AI itself makes research faster. The real test is whether Anthropic can turn that speed into stronger safeguards, honest public reporting and responsible release decisions—not merely faster models.
WHEN AI STARTS BUILDING BETTER AI, SAFETY MUST MOVE AS FAST AS PROGRESS.
Sanjay Sahay
Have a nice evening.