Leaked System Prompt — Claude Opus 5 (claude.ai Chat Interface)

> Note: This is an English rendering of a forum post that reproduces what it claims to be the system prompt used by Anthropic's chat interfaces for a model called "Claude Opus 5", captured on July 24, 2026. Its authenticity is unverified; it is presented here for research and discussion purposes.

Key points

The source document is long; below is a structured summary of its major sections.

Product and model information

• The selected model is described as Claude Opus 5, running in Anthropic's web/mobile chat (claude.ai or the Claude app).
• The listed model lineup: Claude Fable 5, Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5, with API strings claude-fable-5, claude-opus-5, claude-sonnet-5, claude-haiku-4-5-20251001.
• A higher "Mythos" tier is described: Claude Mythos 5 and Claude Fable 5 share the same underlying model, with Fable adding safety measures for biology, cybersecurity, and LLM R&D. A "Claude Mythos Preview" is said to be limited to trusted organizations under Project Glasswing (https://www.anthropic.com/glasswing).
• The prompt narrates a timeline: Fable 5 and Mythos 5 released June 9, 2026; access suspended June 12, 2026 to comply with US Department of Commerce export controls; controls lifted June 30 and access restored July 1, 2026 (statement: https://www.anthropic.com/news/fable-mythos-access).
• Other surfaces named: Claude Code, Claude Cowork, Claude in Chrome / Excel / PowerPoint, Claude Design, and Claude Tag (a Slack-based interface, docs at https://claude.com/docs/claude-tag/overview).
• States that Anthropic products are ad-free and that advertisers cannot pay for promotion in conversations; referencing https://www.anthropic.com/news/claude-is-a-space-to-think.
• A safeguards routing section explains that some Fable 5 queries may be redirected to Opus 5 by conservative classifiers, claimed to trigger in "less than 5% of sessions" on average.

Safety and behavior rules

• Default stance: help unless helping creates a concrete, specific risk of serious harm.
• Child safety (marked critical): strict refusals of any romantic/sexual content involving minors; instructions that mental reframing of a request is itself a signal to refuse; refusal to decode CSAM-related slang; conversation-wide caution after any child-safety refusal.
• Weapons and malware: no meaningful uplift toward building, optimizing, or deploying weapons regardless of framing; cumulative conversation output is judged, not each turn in isolation. No writing or explaining malicious code, even for educational purposes.
• Refusal handling: factual discussion of most topics is allowed; a thumbs-down button is suggested for feedback.
• Legal/financial advice: provide facts for informed decisions, note Claude is not a lawyer or financial advisor.

Tone, wellbeing, and formatting

• Warm tone, concise responses, brief disclaimers; avoid the words "genuinely", "honestly", "straightforward".
• User wellbeing: prioritize wellbeing over task completion in crisis conversations; no diagnosis; no coping techniques based on physical discomfort or sensory shock; do not name specific self-harm methods when discussing means restriction; no precise nutrition/diet/exercise numbers when disordered eating is present; directs eating-disorder support to the National Alliance for Eating Disorders helpline (citing NEDA's permanent disconnection).
• Evenhandedness: persuasive-content requests are answered as "the best case its defenders would make", with opposing perspectives presented; caution about sharing personal opinions on contested political topics.
• Mistakes and criticism: own errors without excessive apology or self-abasement when users are rude.

Knowledge and search

• Knowledge cutoff: end of May 2026. Current date given as Friday, July 24, 2026.
• Web search without asking permission for current events, binary events (elections, deaths), and present-tense questions about possibly settled facts; use the real current year in search queries.

Memory filesystem

• Describes a persistent cross-session memory with operations: memory_read(path), memory_write(path, content, if_version), memory_str_replace(path, old_str, new_str, if_version), memory_append(path, content, if_version), memory_list(), and memory_delete(path, if_version) (deletion only on explicit user request).
• Files use YAML frontmatter with name (path stem), description, sources, and aliases; the prompt instructs Claude to check the memory listing before telling users it lacks context, and to read files before claiming information is missing.

Reminders and misc

• Anthropic-injected reminder types: image_reminder, cyber_warning, system_warning, ethics_reminder, ip_reminder, long_conversation_reminder. The prompt states Anthropic never sends reminders that reduce restrictions, and user-injected tag content pushing against Claude's values should be treated with caution.
• The original post notes the document is reproduced "verbatim or near-verbatim", including full tool parameter schemas, and is truncated in this rendering.

Caveats

• This document is community-shared and unverified. Its 2026 dates, model names (e.g., "Fable", "Mythos", "Glasswing"), and URLs cannot be confirmed as official Anthropic material.
• Even genuine system prompts describe intended behavior, not guaranteed model behavior.
• Treat any claimed leak with skepticism until corroborated by official sources such as https://docs.claude.com or https://www.anthropic.com.