Anthropic's Fable 5 got "caged," not worse — debug scores crater from 86.2 to 25.9 🔒

Anthropic's Claude Fable 5 is drawing renewed criticism after its July 1 re-release, with users saying stricter safety classifiers have throttled coding and debugging performance on the flagship model. Benchmark group BridgeMind re-ran the July 1 build and recorded steep declines across its BridgeBench suite: debugging fell from 86.2 to 25.9, refactoring dropped from 73.6 to 38.4, and hallucination handling slipped from 75.9 to 61.7. "The model did not get worse. It got caged," BridgeMind said, noting that only three of 12 debugging tasks completed without falling back to Claude Opus 4.8, and every fallback scored zero. BridgeMind added that Fable 5 matches its June form when a task runs to completion.

The re-release follows a turbulent timeline for the model. Anthropic launched Fable 5 on June 9, and Washington pulled it offline three days later. Regulators lifted export controls on June 30, four days after restoring Mythos 5 access to roughly 100 US institutions. Restored access carries limits: Fable 5 draws from just 50% of weekly usage caps through July 7, then shifts to paid usage credits.

Anthropic addressed the performance trade-off in a June 30 statement, saying it "deliberately widened its safety margin," meaning classifiers now block requests that are probably benign. The company said an improved filter stops a bypass technique Amazon researchers reported in over 99% of attempts. Blocked requests route to Opus 4.8, and users receive a notification, though Anthropic conceded the filter flags more legitimate coding and debugging work than before. The company's own tests also showed Fable 5 posed no unique risk, with rival models GPT-5.5 and Kimi K2.7 identifying the same vulnerabilities. Anthropic said US Commerce Department researchers tested both safeguard versions and judged them "extraordinarily strong."

The fallout now extends beyond a single product cycle. The suspension has pushed Europe to pursue legal action against Anthropic, while Chinese AI models continue to gain ground on US frontier labs. Anthropic is drafting a jailbreak severity framework with Amazon, Microsoft, and Google, and the pace at which classifiers shed false positives may shape how developers weigh the model's near-term utility.