Independent benchmark shows big drops on Claude Fable 5 after its relaunch, here’s the actual context

Saw this chart from BridgeMind going around. They reran BridgeBench (a coding benchmark covering debugging, refactoring, and hallucination detection) comparing the July 1 relaunch of Fable 5 to the original June 12 version:

Debugging: 86.2 → 25.9
Refactoring: 73.6 → 38.4
Hallucination: 75.9 → 61.7

Some context worth having before jumping to conclusions:

Fable 5 and Mythos 5 got pulled on June 12 due to a Commerce Department export control order, tied to a reported jailbreak that got the model to expose exploitable vulnerabilities. When it came back on July 1, Anthropic added a new safety classifier that catches the reported technique in 99%+ of cases, and any flagged request gets silently rerouted to Opus 4.8 instead of refused outright.

That’s the mechanism BridgeMind is pointing at. Their claim isn’t that the underlying weights changed, it’s that the classifier is triggering on too many normal coding tasks and quietly downgrading people to Opus 4.8 without them realizing it. A few other users on X are reporting the same thing (constant fallback, slower one-shot performance).

No independent lab has confirmed whether the weights themselves changed. This might just be an overly aggressive classifier rather than an actual capability regression, but if you’re relying on Fable 5 for coding work, worth watching this closely before you assume you’re getting the same model you had before June 12.