Why Now? From Mythos Preview to Fable 5
Claude Mythos Preview — Anthropic's previous frontier model — was powerful enough to identify thousands of cybersecurity vulnerabilities. Rather than risk that capability being misused, Anthropic restricted it to Project Glasswing: a US government partnership focused on finding and patching critical infrastructure flaws before bad actors could exploit them.
Over the following months, Anthropic engineered automated safety classifiers that detect high-risk requests (cybersecurity exploits, dangerous biology) and redirect them to Claude Opus 4.8. Those classifiers are now robust enough that the underlying Mythos-class capability can be released broadly — that release is Claude Fable 5.
Claude Mythos 5 also launched simultaneously: it's the same model as Fable 5 but with the cyber and biology classifiers lifted, available only to Glasswing partners and approved biosecurity researchers.
Performance at a Glance
Fable 5 vs. Mythos Preview
| Feature | Claude Mythos Preview | Claude Fable 5 |
|---|---|---|
| Access | Glasswing partners only | Generally available |
| Pricing (in/out) | Higher | $10 / $50 per M tokens |
| Cyber safety classifiers | None | Auto-classify & redirect |
| Long-horizon agents | Yes | Enhanced |
| Biology safety classifiers | None | High-risk requests blocked |
Fable 5 is fully available on the Claude API and consumption-based Enterprise plans from launch day. For Pro, Max, Team, and seat-based Enterprise subscription plans, Fable 5 is included at no extra cost through June 22. From June 23 onward, using it will require usage credits. Anthropic plans to restore Fable 5 as a standard subscription inclusion once capacity allows.
Built for Long-Horizon Agentic Work
The defining capability of Fable 5 is sustained autonomous operation. While earlier Claude models excel at individual tasks, Fable 5 is designed to maintain coherent, goal-directed work across multi-day projects — coding agents, research pipelines, financial analysis workflows, and legal review tasks that previously required constant human supervision.
Key supported capabilities:
- Effort control and Task Budgets (manage inference cost per task)
- Memory Tool — persistent context across sessions
- Code Execution — sandboxed runtime
- Context Editing (beta) — trim and restructure long conversations
- Compaction — efficient long-context management
Use model ID
claude-fable-5 on the Claude API, Amazon Bedrock, Google Vertex AI, or Microsoft Foundry. Mythos 5 uses claude-mythos-5 and requires Glasswing access. Pricing is identical for both: $10 input / $50 output per million tokens, with a 90% discount on cached input tokens.Safety Architecture: The Classifier System
Rather than wholesale restricting a model's capabilities, Anthropic built domain-specific safety classifiers that intercept requests touching high-risk areas — cyberweapon development, dangerous pathogens — and route them to Claude Opus 4.8 instead. Fable 5 itself handles everything else. Anthropic acknowledges the classifiers are broad today and commits to refining them over time so legitimate security research and biology queries are handled appropriately.
This architecture represents a shift in how frontier AI labs think about safety: instead of choosing between capability and safety, Anthropic is betting on fine-grained real-time intervention.
Key Takeaways
- Claude Fable 5 is the first Anthropic model bringing Mythos-class capability to the general public
- 95% on SWE-bench Verified, 80% on SWE-bench Pro
- $10 / $50 per million tokens — more than 50% cheaper than Mythos Preview
- Designed for multi-day autonomous agentic workflows without human intervention
- Automatic safety classifiers redirect high-risk requests to Claude Opus 4.8
- Claude Mythos 5 (classifiers lifted) is available simultaneously for Glasswing partners
— Anthropic Official Announcement — Claude Fable 5 & Mythos 5
— Claude Fable 5 Product Page (Pricing & Availability)
— Developer Docs — API Integration & Model IDs