Claude Fable 5 came back online on July 1, 2026, and BridgeBench AI tested that July 1 version. BridgeBench reported Debugging fell from 86.2 to 25.9, Refactoring from 73.6 to 38.4, and Hallucination resistance from 75.9 to 61.7. The episode is noteworthy because many users perceived Fable 5 as ‘nerfed’ by a gatekeeper/router after a safety classifier deployed as a condition of Fable’s reinstatement became more aggressive and rerouted numerous tasks to Claude Opus 4.8.
Anthropic’s safety classifier was deployed as a condition of Claude Fable 5’s reinstatement. The classifier was trained to block the Amazon-reported jailbreak technique that had previously allowed Fable 5 to identify and demonstrate software vulnerabilities. The gatekeeper in front of Fable 5 became much more aggressive after the classifier’s deployment. The July 1 version of Claude Fable 5 was the basis for the tests that evaluated routing and model output.
During BridgeBench AI testing of TypeScript debugging tasks, only three of 12 tasks reached Fable 5 while nine were intercepted by Anthropic’s safety classifier and rerouted to Claude Opus 4.8. BridgeBench reports that it assigns a score of zero to the Opus fallback for those intercepted tasks. BridgeBench’s aggregated scores for the July 1 tests declined in multiple categories, including debugging, refactoring, and hallucination resistance. The classifier intercepted and rerouted a substantial number of test cases, changing which model produced the final outputs in those cases.
Anthropic’s safety classifier was deployed as a condition of Claude Fable 5’s reinstatement. The classifier was trained to block the Amazon-reported jailbreak technique that had previously allowed Fable 5 to identify and demonstrate software vulnerabilities. After the classifier’s deployment, the gatekeeper in front of Fable 5 became much more aggressive. The July 1 version of Claude Fable 5 was the basis for the tests that evaluated routing and model output.
BridgeBench AI testing of TypeScript debugging tasks found that only three of 12 tasks reached Fable 5 while nine were intercepted by the safety classifier and rerouted to Claude Opus 4.8. BridgeBench reports that it assigns a score of zero to the Opus fallback for those intercepted tasks. BridgeBench’s aggregated scores for the July 1 tests declined across multiple categories, with debugging reported as 86.2 → 25.9, refactoring as 73.6 → 38.4, and hallucination resistance as 75.9 → 61.7. The intercepted cases therefore produced final outputs from the fallback model rather than from Fable 5.
Arena AI’s benchmarking methodology uses Elo scoring across thousands of blind human-preference votes in head-to-head matchups across multiple categories. The safety classifier’s deployment and the resulting rerouting were part of the reinstatement conditions evaluated in post-reinstatement testing. The reported routing outcomes and scores reflect results from the July 1 tests.
The model itself did not get dumber; the gatekeeper in front of Claude Fable 5 became more aggressive and restrictive following reinstatement. This change in the router’s behavior coincided with deployment of a safety classifier as a condition of Fable’s reinstatement. “The catch is in the methodology.” Observers characterized the effect as an impression of “nerfing” caused by routing decisions rather than a reduction in the underlying model’s capability.
The new guardrails have been reported to intercept and reroute a substantial number of tasks away from Fable 5 to Claude Opus 4.8. BridgeBench testing of 12 TypeScript debugging tasks found that nine were intercepted by the safety classifier and rerouted to Opus while three reached Fable 5. BridgeBench assigns a score of zero to the Opus fallback for those intercepted tasks. “The new guardrails are kicking in on way too many tasks and falling back to Opus…”
Claude Fable 5 came back online on July 1, and the reinstatement included a deployed safety classifier and a more aggressive gatekeeper/router operating in front of the model. That enforcement-layer behavior rerouted a substantial share of tasks to Claude Opus 4.8 and produced an impression that Fable 5 had been “nerfed,” distinguishing the underlying model capability from the effects of enforcement mechanisms.
