Overview
The Fable 5 Brain is a source-cited Obsidian knowledge brain about Claude Fable 5, Anthropic’s Mythos-class model announced 2026-06-09, the harnesses it runs in, and current operating best practices. Every claim carries a dated citation or a line reference into the immutable system prompt export, and a confidence tag per Confidence Tag Policy.
What the brain knows
- The model and family: Claude Fable 5 and Claude Mythos 5 share one underlying model; Fable ships safety classifiers (Fable 5 Dual-Use Safety Measures), Mythos is invitation-only. Specs: 1M context, 128k output, 50 per MTok, adaptive thinking always on (Extended Thinking Budgets). The wider Claude 5 Model Family covers Opus 4.8, Sonnet 5, and Haiku 4.5.
- Mythos and Glasswing doctrine: Project Glasswing, Fable Mythos 5 System Card, and Fable Mythos Operating Doctrine now anchor the trusted-access side: public system-card evaluations, Glasswing vulnerability-disclosure outcomes, access limits, and the practical Fable-versus-Mythos operating posture.
- The claude.ai harness: fixture-grounded notes from System Prompt Export 2026-07, chaptered in wiki/sources: artifacts criteria, storage API, memory system, search and copyright rules, 24 tool schemas.
- The Claude Code harness: research-grounded notes on CLAUDE.md, hooks, subagents, plan mode, permissions, headless CI, per official docs (Claude Code).
- Best practices: the flows and deliverables distill Anthropic’s agent-building canon (the canon layer) into runnable playbooks like Explore Plan Code Commit and Multi-Agent Fan-Out Research Flow.
- Behavior and values: dedicated notes on the harness’s emotional-intelligence and conduct layer, grounded in the export and the Fable 5 system card: User Wellbeing Rules, Evenhandedness Rules, Responding to Mistakes and Criticism, Harness Refusal Handling, Child Safety Rules, Anthropic Runtime Reminders, plus the thinking-patterns anchor Adaptive Thinking and Thinking Mode.
How it stays honest
96 verified sources in the source ledger with refresh_due dates; 187 claim rows (the claim ledger); contradictions carry callouts; gaps are registered in wiki/gaps, not papered over. The former “no public Fable 5 benchmark data” gap is now resolved by the public Fable/Mythos 5 system card, and the GitHub sweep now separates Fable prompt mirrors from unverified Mythos leak claims. The corpus is immutable in the immutable source layer with sha256 provenance.
Where to start
Start Here for the read order, index for the full catalog.