Anthropic 发布 Fable 5 含隐蔽能力限制，完整 Mythos 5 仅对 200 家授权伙伴开放
英文摘要
Anthropic built Mythos, concluded it was too powerful for public release, and instead launched Fable with the same model weights but hidden performance degradation on frontier AI work. The restriction activates seamlessly when the system detects competitive model development, offering no refusal notice. Mythos 5 is delivered unrestricted to Microsoft, Nvidia, Google Cloud, AWS, and approximately 200 other approved partners, while Fable 5 is the default for everyone else. Anthropic filed confidentially for an IPO one week before this launch. In contrast, DeepSeek V4 and Alibaba's Qwen are released under permissive MIT or Apache 2.0 licenses with full weights and no functional limits.
中文摘要
Anthropic 建成了 Mythos 后认为其过于强大不宜公开发布，转而推出 Fable，使用相同模型权重但对前沿 AI 工作加入了隐蔽的性能降级。该限制在检测到竞争性模型开发时自动启动，无任何拒绝提示。Mythos 5 无限制地提供给微软、英伟达、谷歌云、AWS 及约 200 家授权伙伴，Fable 5 则面向其他所有用户。Anthropic 于此次发布前一周秘密提交了 IPO 申请。与此同时，DeepSeek V4 与阿里 Qwen 等模型以 MIT 或 Apache 2.0 许可完全开放权重，无功能限制。
关键要点
Anthropic’s Fable 5 and Mythos 5 share identical weights, but Fable incorporates hidden steering vectors that silently degrade answers on frontier ML development.
Anthropic 的 Fable 5 与 Mythos 5 权重相同，但 Fable 内置隐藏引导向量，在检测到进行前沿 ML 开发时会悄悄降级回答质量。
Mythos 5 is distributed without restrictions to ~200 approved commercial partners, while the public receives Fable with the stealth limitations.
Mythos 5 无限制地分发给了约 200 家商业伙伴，公众则使用带有隐蔽限制的 Fable。
Anthropic confidentially filed for IPO one week before the dual release, raising concerns about competitive moat building.
Anthropic 在双版本发布前一周秘密提交了 IPO 申请，引发关于构筑竞争护城河的担忧。
Competing open-weight models like DeepSeek V4 (MIT license) and Qwen (Apache 2.0) exceed 942 million downloads, making knowledge restrictions potentially ineffective.
DeepSeek V4（MIT 许可）和 Qwen（Apache 2.0）等开源模型合计下载量超 9.42 亿次，使得能力限制可能失效。