On September 2, 2026, Anthropic launched the large-model version Fable 5.1, which thoroughly blocks illegal distillation by strictly verifying the context consistency of “thinking blocks” in API calls. The new rule requires that the thinking blocks returned by clients must be completely consistent with system prompts, tools, and historical messages; otherwise, service will be denied. A non-strict mode is also provided to maintain dialogue continuity. This mechanism applies only to new accounts created after August 31, 2026; existing accounts and C-end users are not affected. Additionally, restrictions were added to the prompts in Claude’s consumer application system: retelling song lyrics, poetry, or book excerpts, as well as creating copyrighted characters and artworks, are prohibited. For such requests, the model will deny them and proactively offer alternative descriptions or analyses.
AnthropicClaudeFable 5.1Haiku 4.5Sony Music Publishing
Event frameEVENT FRAME
Launch
科技 · AIacross 2 days
Status
Anthropic released Fable 5.1 and updated Claude system prompts, implementing strict anti-distillation and copyright restrictions.
Coverage · reports per dayLANGUAGE SPLIT
Coverage mixSOURCE MIX
Chinese media 1English media 1
Entity relations
Integrated timelineUNIFIED TIMELINE
2026-09-02
Simon Willison reports on Claude copyright restrictions
Newly added strict restrictions in Claude system prompts: prohibiting repetition of song lyrics, poetry, book excerpts, and drawing copyrighted characters.
2026-09-05
Security internal report on new anti-distillation rules in Fable 5.1
Anthropic launched Fable 5.1 on September 2, blocking illegal distillation by verifying the consistency of thought block contexts.
Anthropic has released system prompts for its Claude consumer applications (Claude.ai and mobile versions), covering both current versions and historical changes. The new version of prompts includes strict restrictions: retelling lyrics from songs, poems, or book excerpts is prohibited; drawing copyrighted characters, logos, or specific artworks is also forbidden. For requests related to lyrics, Claude will refuse and provide alternative descriptions or analyses; for copyrighted characters (such as Sonic), even if the pose or style is adjusted, the request will still be rejected if the overall image is recognizable. The platform’s documentation supports Markdown format export, making it easy to compare differences between different versions of the prompts.
On September 2, Anthropic launched Fable 5.1, which strictly verifies the consistency of the “thinking blocks” in API calls, completely blocking the path for illegal distillation of large models. This mechanism requires that the thinking blocks returned by clients must be exactly consistent with the system prompts, tools, and historical messages that generated them; any alteration will result in a service rejection. Additionally, a “non-strict mode” is available, allowing context modifications while automatically deleting thinking blocks to maintain dialogue continuity. This update aims to curb the use of fake accounts and automated scripts to大规模 steal model reasoning logic, preventing the risk of AI becoming uncontrollable due to the “disconnection between capabilities and security”. The new rules apply only to new API accounts created after August 31, 2026; existing old accounts and C-end users are not affected. Legal developers can migrate by maintaining context consistency or enabling the non-strict mode, which will also significantly reduce call costs and shorten response times by improving prompt cache hit rates.