Anthropic found that small models can perform individual replication tasks but remain unreliable across the complete sequence. Stronger models succeeded in simplified evaluations.

Its end-of-year capability estimate depends on continued progress. The laboratory setup does not establish reliable propagation across the open internet.

Go to the source

This is an original summary, not a republication of the full report.

Read the original at Reuters ↗

Primary statement: Anthropic research, 12 October 2026 ↗