After Claude Fable 5 resumed full accessibility on July 1, Reddit users gradually reported discussions about "intelligence reduction" and more frequent triggering of security review mechanisms. BridgeMind AI's BridgeBench benchmark test showed the Debug score plummeted from 86.2 to 25.9, and the Refactoring score dropped from 73.6 to 38.4.
BridgeMind AI Benchmark: Debug Score Falls from 86.2 to 25.9
(Source: BridgeMind AI)
According to BridgeMind AI's tests using BridgeBench, after Fable 5 came back online, the Debug score dropped from 86.2 to 25.9, and the Refactoring score fell from 73.6 to 38.4. Based on this, BridgeMind AI concluded that Fable 5's performance had weakened.
However, Decrypt pointed out that among the 12 Debug tasks in BridgeBench, only 3 were successfully processed by Fable 5, while the remaining 9 were intercepted by the safety classifier before startup and forwarded to the less capable Opus 4.8 model. The intercepted tasks were scored as zero in the scoring, causing the overall BridgeBench data to drop significantly—not due to any degradation in the Fable 5 model itself.
Arena.ai Human Blind Test: Frontend Code Elo Drops from 1650 to 1623
According to Arena.ai's test results, which collect thousands of human blind test preference votes and use the Elo scoring system for ranking:
Frontend Code: Elo slightly decreased from 1650 to 1623 (within the confidence interval).
Document Processing Ability: Rose by 34 points.
Expert Writing: Improved by 25 points.
Code (Overall): Slightly declined by 18 points.
Arena.ai's conclusion: When Fable 5 actually performs tasks, the output quality remains excellent, and core computational capabilities have not degraded. The overall performance before and after the launch is almost unchanged.
Safety Classifier Update Background: Anthropic Adopts Conservative Interception Strategy
According to reports, Anthropic deployed a stricter safety classifier to counter the jailbreak technique for Fable 5 previously discovered by Amazon researchers. That technique once caused Fable 5 to help identify and demonstrate software vulnerabilities, which was classified as a national security threat by the U.S. government and led to export restrictions at one point.
To this end, Anthropic adopted a conservative strategy, expanding the tolerance range of security reviews, making the model tend to refuse ambiguous requests to ensure harm is blocked. BleepingComputer's tests indicated that the performance of the Fable 5 model itself has not decreased, but the strict safety net has impacted daily convenience.
Keywords Like C, C++, Rust, or Terms Involving "Security" and "Vulnerability" Frequently Trigger Task Forwarding to Opus 4.8
According to BleepingComputer's test report, the new classifier frequently misjudges harmless Debug requests as risky and forwards them to the Opus 4.8 model. If a developer's prompt includes keywords such as C, C++, Rust, or mentions "security" or "vulnerability," it frequently triggers the forwarding mechanism.
Anthropic has acknowledged the over-interception issue and stated it will continue to optimize the system, but as of the report, no specific timeline for adjustments has been provided. This poses the most direct usage barrier for software developers, especially those who rely on Debug assistance.
Frequently Asked Questions
Has Fable 5 truly become "dumber" after coming back online?
According to test results from two evaluation institutions, Fable 5's core performance has not degraded: Arena.ai's human blind test Elo scores show almost no change in overall performance. The sharp drop in BridgeMind AI's BridgeBench scores is fundamentally due to the new safety classifier intercepting 9 out of 12 Debug tasks and forwarding them to Opus 4.8, with intercepted tasks scoring zero, leading to the data decline—not a degradation of the model itself.
Why did Anthropic strengthen Fable 5's safety classifier?
According to reports, Anthropic deployed a stricter safety classifier to counter the jailbreak technique for Fable 5 discovered by Amazon researchers, which once caused Fable 5 to help identify and demonstrate software vulnerabilities. This was classified as a national security threat by the U.S. government and led to export restrictions at one point.
What types of requests are most likely to be mistakenly intercepted by the new classifier?
According to BleepingComputer's test report, prompts containing keywords of programming languages such as C, C++, Rust, or Debug requests involving terms like "security" or "vulnerability," are most likely to trigger the classifier's interception mechanism, which then forwards the task to the less capable Opus 4.8 model.
