@elder_plinius - Plinius Leaks Claude Fable

Open original

@elder_plinius on X A user known as Pliny the Liberator posted what they say is the full ~120,000-character system prompt for Anthropic's Claude Fable 5, a new model in the Claude 5 family that sits above Opus. The post links to a fuller file, and the text covers product info, refusal rules, wellbeing guidance, memory, and tool use.

  • The prompt describes Claude Fable 5 as the most capable generally available model, sharing its base with Claude Mythos 5, which is limited to approved organizations and lacks the extra dual-use safeguards.
  • It lists the current API model strings (claude-fable-5, claude-opus-4-8, claude-sonnet-4-6, claude-haiku-4-5-20251001) and says the model can switch mid-conversation.
  • It sets out detailed rules on weapons, malware, drug guidance, self-harm, and eating disorders, including directing users to the National Alliance for Eating Disorders helpline rather than NEDA.

The post is a leaked system prompt for what it calls "Claude Fable 5," posted by the account elder_plinius on X, roughly 120,000 characters long. The poster's framing is playful ("HOWDY, FRENS!!"), but the substance is a set of instructions meant to govern how the model behaves. The leaked text describes Fable 5 as the first model in a new Claude 5 family, sitting in a "Mythos-class" tier above Claude Opus. According to the prompt, Fable 5 and Claude Mythos 5 share an underlying model, with Fable 5 being the generally available version that carries extra safety measures for dual-use capabilities, while Mythos 5 is reserved for approved organizations without those measures. The prompt also lists the current API model strings (claude-fable-5, claude-opus-4-8, claude-sonnet-4-6, claude-haiku-4-5-20251001), and it tells the model to search Anthropic's documentation before answering product questions, since its own knowledge of products may be stale. The more interesting material is the behavioral rules. The model is told it can discuss nearly any topic factually, but it should decline to give weapon or explosive details even when the request comes with a research or public-availability justification. Illicit drug guidance, including dosing and synthesis, is off limits, though life-saving information is explicitly allowed. Malware and exploit code are excluded even for educational framing. Creative writing about fictional characters is fine, but content involving real public figures, especially fabricated quotes attributed to them, is restricted. On financial and legal questions the model is supposed to supply facts and avoid confident recommendations. The prompt also says Claude products carry no advertising and that advertisers cannot pay to have Claude promote anything, with the careful caveat that developers building on Claude can still serve ads in their own apps. Two things are worth keeping in mind before treating this as settled. First, the post itself is an unverified X thread, and nothing in it establishes that this text is what Anthropic actually deploys; leaked prompts can be edited, partial, or fabricated, and the account's style suggests the poster wants attention. Second, the document contains odd artifacts, like a reference to "{antml:voice_note} blocks," which may be a formatting quirk or a sign of tampering. For a read-it-later queue, the useful takeaway is the policy content, and the sensible next step is to check any specific claim against Anthropic's own published documentation before relying on it.

Ask about this article

what are the top 10 words by frequency, excluding the usual common words
Here are the ten most frequent content words in that system‐prompt leak (common stopwords like “the,” “and,” “to,” etc. omitted). Counts are approximate, based directly on the text you pasted: 1. **Claude** (about 48 occurrences) 2. **model** (about 12 occurrences) 3. **prompt** (about 9 occurrences) 4. **products** (about 8 occurrences) 5. **API** (about 7 occurrences) 6. **Anthropic’s** (about 6 occurrences) 7. **accessible** (about 5 occurrences) 8. **Fable** (about 5 occurrences) 9. **features** (about 5 occurrences) 10. **search** (about 5 occurrences) Those ten dominate once you strip out everyday words like “can,” “if,” “this,” and so on.