As the chart implies, the issue isn’t that the underlying weights of Fable 5 were made fundamentally “dumber”—it’s that Anthropic wrapped the model in an incredibly aggressive, restrictive policy routing architecture to satisfy compliance demands.

According to Anthropic’s documentation and recent technical breakdowns, the “nerf” comes down to a few major bottlenecks:

As developers quickly noticed—and as the BridgeBench data in HMOulJbWYAAw-bf.jpg proves—this “fix” came at a devastating cost to the model’s actual utility, cratering debugging scores by over 60 points.

For 19 days the best model on earth was illegal to show a foreign national, including Anthropic’s own staff. Yesterday Fable 5 came back, with an admitted new classifier, a silent reroute to Opus 4.8, and no proof the weights are the same. Nobody can publish that proof, and the vendor didn’t try. Access used to be gated by price. Now it’s gated by permission, and verified by vibes.