Opus 5.5 becomes the first Opus to restrict frontier large model development: triggers downgrade to Opus 5

动察 Beating AI News Flash: For the first time, Anthropic has brought the "frontier model R&D" restrictions from the Fable series into Opus. Opus 5.5 has added a classifier similar to Fable's. When encountering a small number of frontier LLM R&D tasks, such as developing kernels for specific AI accelerators, the system will directly switch from Opus 5.5 to the less capable Opus 5. Opus 5 does not have this restriction.

Fable 5 previously already restricted distributed training infrastructure, AI accelerator design, and some kernel development. Anthropic explained at the time that strong models may already accelerate next-generation model R&D, and it does not want Claude to help competitors build frontier models of the same level faster. Fable currently also blocks or switches models for a small subset of such tasks.

Anthropic's own evaluations also run with this set of guardrails enabled. Once a frontier LLM R&D question triggers the restriction, it is answered by Opus 5 instead, and Opus 5.5's related benchmark scores may also be dragged down as a result.

Anthropic emphasizes that this type of classifier covers only a very small portion of frontier LLM R&D tasks, and the vast majority of ordinary AI, machine learning research, and everyday programming will not be affected. After the downgrade is triggered, Claude will clearly show that it has switched to Opus 5, and subsequent conversations will continue to use Opus 5.