Claude Fable 5 Backlash Grows
摘要
文章称 Anthropic 的 Claude Fable 5 于 6 月 9 日推出后因出口管制下线，6 月 30 日恢复访问并于 7 月 1 日重新发布；BridgeMind 的 BridgeBench 测试显示调试分数从 86.2 降至 25.9、refactoring 从 73.6 降至 38.4、幻觉处理从 75.9 降至 61.7，仅 12 个调试任务中 3 个无需 fallback 到 Opus 4.8。Anthropic 表示底层模型不变，是安全分类器更严格以阻挡更多网络安全任务，阻挡请求会转到 Opus 4.8 并通知用户，同时承认误判了更多合法编码调试工作；模型恢复后有每周使用上限，Fable 5 仅占 50% 直至 7 月 7 日。文章还提及 Anthropic 正与 Amazon 等公司制定 jailbreak 严重性框架，以及欧洲和中国模型的竞争背景。
荐读理由
BridgeBench 数据显示 Fable 5 debugging 从 86.2 跌到 25.9、refactoring 从 73.6 跌到 38.4，Anthropic 承认是安全分类器拦截更多 coding 任务而非模型变弱，这给你判断安全 margin 宽化对实际可用性的影响提供了具体案例。
原文
Claude Fable 5 Backlash Grows as Users Say Anthropic ‘Caged’ Its Flagship AI
Thu, July 2, 2026 at 9:30 PM UTC
US Lifts Export Controls on Anthropic's Claude Fable 5 and Mythos 5 Models. Photo by BeInCrypto
Anthropic's Claude Fable 5 faces growing backlash after its July 1 re-release. Users claim stricter guardrails have crippled the flagship model's coding, debugging, and agentic performance.
Benchmark group BridgeMind reported steep score drops across its BridgeBench suite. Meanwhile, Anthropic maintains the underlying model is unchanged and attributes the friction to tighter safety classifiers.
Claude Fable 5 Benchmark Scores Collapse After Re-Release
BridgeMind re-ran the July 1 version of Fable 5 and recorded sharp declines. 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.
BridgeBench scores for Claude Fable 5 before and after the re-release, Source: Users on X
The mechanics behind those numbers matter. Only three of 12 debugging tasks were completed without falling back to Claude Opus 4.8, and every fallback scored zero.
Advertisement
Advertisement
Therefore, the collapse reflects blocked tasks rather than weaker reasoning.
BridgeMind stressed that Fable 5 matches its June form when a task runs to completion.
"The model did not get worse. It got caged," they indicated.
Follow us on X to get the latest news as it happens
The timeline explains the tension. Anthropic launched Fable 5 on June 9, and Washington pulled it offline three days later. Regulators lifted its export controls on June 30, four days after they restored Mythos 5 access for roughly 100 US institutions.
Restored access also carries limits. Fable 5 draws from just 50% of weekly usage caps through July 7, then shifts to paid usage credits.
Anthropic Defends Its Wider Safety Margin
Anthropic addressed the trade-off in a June 30 statement. The company said it deliberately widened its safety margin, meaning classifiers now block requests that are probably benign. An improved filter stops the bypass technique, Amazon researchers reported in over 99% of attempts.
Claude Fable 5 will be available again globally tomorrow.
After a series of productive conversations with the US government, we're redeploying the model with a new set of classifiers to target and block more cybersecurity tasks. In the near term, some routine tasks like coding…
— Anthropic (@AnthropicAI) July 1, 2026
Blocked requests route to Opus 4.8, and users receive a notification. However, Anthropic conceded the filter flags more legitimate coding and debugging work than before.
Advertisement
Advertisement
Its own tests also showed Fable 5 posed no unique risk. Rival models, including GPT-5.5 and Kimi K2.7, identified the same vulnerabilities.
Anthropic says US Commerce Department researchers tested both safeguard versions and judged them extraordinarily strong.
The stakes reach beyond one product cycle. The suspension pushed Europe to court Anthropic, while Chinese AI models gain ground on US frontier labs.
Anthropic is now drafting a jailbreak severity framework with Amazon, Microsoft, and Google. Whether classifiers shed false positives quickly may determine whether power users stay or defect.
Read the Original story Claude Fable 5 Backlash Grows as Users Say Anthropic 'Caged' Its Flagship AI by Lockridge Okoth at beincrypto.com
这条对你有帮助吗？