METR finds progress in Claude Opus 5.5, but not full research automation

METR published its predeployment evaluation of Claude Opus 5.5 on September 22, 2026. It assesses research capabilities, not alignment.

The researchers expect modest gains over Fable 5.1. They consider full automation of AI research unlikely with this model.

Testing covered five tasks over ten business days of API access. Difficult, extended workflows still exposed weaknesses.

Anthropic could review and edit the report; METR approved the final text. Some additional evidence is not public. These findings do not provide a general safety clearance.