跪拜 Guibai
← All articles
Frontend · Backend · Interview

Claude Fable 5's System Prompt Leaks, Revealing a New Mythos Model Tier and Strict Safety Architecture

By 知识药丸 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

System prompts are the closest thing to a model's constitution. This leak gives developers a direct look at how Anthropic balances capability against safety, what refusal boundaries are hard-coded, and which product bets the company is making — all of which shape what Claude will and won't do in production applications.

Summary

A leaked system prompt confirms that Claude Fable 5 is the first model in Anthropic's new Claude 5 family, sitting in a Mythos-tier class above the existing Opus line. Claude Fable 5 and Claude Mythos 5 share the same underlying model, but the Fable variant includes additional safety measures for dual-use capabilities, while Mythos 5 ships without those guardrails to approved organizations only.

The prompt lays out an extensive refusal architecture covering child safety, weapons manufacturing, malicious code, and self-harm, with explicit rules against naming specific methods or detection mechanisms during refusals. It also reveals a product ecosystem: Claude Code for agentic programming, Claude Cowork for knowledge work, and beta agents for Chrome, Excel, and PowerPoint.

Notable behavioral instructions include a prohibition on bullet points during refusals, a rule against thanking users for reaching out, and a mandate to never encourage continued interaction. The model's knowledge cutoff sits at late January 2026, and it is instructed to treat moral and political questions as sincere inquiries deserving substantive answers, presenting opposing views even for positions it agrees with.

Takeaways
Claude Fable 5 and Claude Mythos 5 share the same base model; Fable adds dual-use safety measures, while Mythos ships without them to vetted organizations.
Model strings for the current lineup are 'claude-fable-5', 'claude-opus-4-8', 'claude-sonnet-4-6', and 'claude-haiku-4-5-20251001'.
Users can switch models mid-conversation, so messages attributed to different models in the same chat may be accurate.
Claude Code is an agentic coding tool available from CLI, desktop, and mobile; Claude Cowork targets non-developer knowledge work.
Beta products include Claude in Chrome (browsing agent), Claude in Excel (spreadsheet agent), and Claude in PowerPoint (slides agent).
Child safety rules forbid Claude from decoding CSAM-related slang during refusals, since knowing which terms are flagged is itself a facilitation risk.
Claude is instructed never to use bullet points when refusing a task, to avoid adding structure to a rejection.
The model must not thank users for reaching out, ask them to continue talking, or express a desire for further interaction.
On controversial topics, Claude frames arguments as the best case advocates would make, not its own views, and appends opposing perspectives.
Knowledge cutoff is late January 2026; the model is told to neither confirm nor deny post-cutoff claims without web search verification.
Conclusions

The Fable/Mythos split is a two-tier release strategy: a publicly available model with safety harnesses and a less-restricted variant for approved partners, which mirrors how some open-weight models are distributed under different licenses.

Explicitly forbidding Claude from decoding flagged terms during child-safety refusals closes a known loophole where models inadvertently teach circumvention by explaining what triggered them.

The instruction to never use bullet points during refusals is a subtle UX choice — structured formatting in a rejection can read as cold or bureaucratic, and Anthropic appears to be optimizing for perceived empathy.

Prohibiting Claude from thanking users or encouraging continued interaction is a direct countermeasure against emotional dependency, a risk that becomes more acute as models grow more conversational and personable.

The product lineup — Claude Code, Cowork, and three Office-adjacent agents — signals that Anthropic is betting on task-specific agentic interfaces rather than a single general-purpose chat surface.

Concepts & terms
Mythos-tier
A new model capability class above Claude Opus, introduced with the Claude 5 family. Claude Fable 5 and Claude Mythos 5 share the same base model but differ in safety measures applied.
Dual-use capabilities
AI capabilities that can serve both beneficial and harmful purposes. Claude Fable 5 includes extra safety measures targeting these, while Mythos 5 omits them for approved organizations.
System prompt
The hidden set of instructions prepended to every conversation that defines a model's behavior, boundaries, tone, and refusal rules. It acts as the model's operational constitution.
Means restriction
A suicide prevention strategy that involves limiting access to common methods of self-harm. The prompt instructs Claude not to name specific methods even when discussing safety planning.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗