What Is Claude Fable 5?
On June 9, 2026, Anthropic released Claude Fable 5, making the Mythos architecture publicly available for the first time. Previously, this same underlying architecture (Claude Mythos Preview) had been restricted to vetted cybersecurity defenders through Project Glasswing. Fable 5 represents Anthropic's answer to the challenge of deploying frontier-level intelligence both safely and broadly.
Why Mythos Preview Was Never Released Publicly
During training and evaluation of Mythos Preview, Anthropic discovered the model was capable of independently finding thousands of cybersecurity vulnerabilities. Rather than risk misuse, Anthropic chose not to release it publicly — instead deploying it to help defenders patch critical infrastructure flaws before bad actors could exploit them. Fable 5 is the result of that safety-first detour: a model with the same raw capability, refined for responsible broad deployment.
How the Safety Guardrails Work
Fable 5 introduces an automated safety routing system. Queries touching high-risk domains — cybersecurity, biology, chemistry — are automatically reviewed and, if flagged, redirected to Opus 4.8 instead. This means users still get a highly capable response; it just comes from a model without Fable 5's frontier-level capabilities in those specific areas.
Anthropic reports the guardrails trigger in fewer than 5% of sessions on average, and acknowledges they are currently calibrated conservatively — some benign requests may be caught. The team is actively working to reduce false positives as they build more capacity.
From launch through June 22, 2026, Pro, Max, Team, and seat-based Enterprise subscribers can use Fable 5 at no extra cost. After June 22, usage credits will be required. Anthropic has indicated they intend to restore Fable 5 as a standard plan feature once capacity allows.
Claude Fable 5 vs. Claude Mythos 5
Alongside Fable 5, Anthropic simultaneously launched Claude Mythos 5 — the same underlying model but with safeguards lifted in certain areas. Mythos 5 is currently restricted to Glasswing partners (government cyberdefenders and critical infrastructure providers) with the cyber safeguards removed. A broader trusted access program for select biology researchers and others is planned for the near future.
Both Fable 5 and Mythos 5 are priced at $10/M input and $50/M output — more than 50% cheaper than Mythos Preview. Performance is comparable to or slightly better than Mythos Preview in most benchmarks.
What Fable 5 Is Built For
Anthropic emphasizes Fable 5's ability to sustain autonomous effort on complex, long-horizon tasks — potentially operating for days without human intervention. Target use cases include:
- Software engineering: large-scale refactoring, multi-repo debugging, extended agentic coding sessions
- Research & analysis: deep literature review, multi-source synthesis, financial modeling
- Legal & compliance: document review, contract analysis, regulatory mapping
- Science: experimental design, data interpretation, literature summarization
Key Takeaways
- Claude Fable 5 is the first publicly available Mythos-class model (launched June 9, 2026)
- Safety guardrails auto-route risky queries to Opus 4.8 — affects fewer than 5% of sessions
- API pricing: $10/M input, $50/M output — over 50% cheaper than Mythos Preview
- Free on paid subscription plans through June 22; usage credits required after that
- Claude Mythos 5 released simultaneously but restricted to Project Glasswing partners
- Available on Claude API, AWS, Google Cloud, and Microsoft Foundry
Availability & Access
| Plan | Availability | Notes |
|---|---|---|
| Claude API | Immediate, full | claude-fable-5 model ID |
| Consumption-based Enterprise | Immediate, full | No extra setup |
| Pro / Max / Team / Enterprise | Now through June 22 | Included at no extra cost |
| Pro / Max / Team / Enterprise | After June 22 | Usage credits required |
| AWS / GCP / Microsoft Foundry | Immediate | Full deployment |
Infrastructure Scale
The Fable 5 launch is backed by a $35 billion AI compute expansion financed by Apollo and Blackstone via Broadcom's XPV platform. The initiative targets over 1 gigawatt of AI computing capacity from mid-2026, with a broader goal of 20+ gigawatts through 2028 — enough to sustain frontier model development at scale across multiple labs.
— Anthropic Python SDK (GitHub) — Official library for Claude API integration
— Claude Code (GitHub) — Agentic coding CLI powered by Fable 5
— Anthropic Cookbook (GitHub) — Prompt examples and API usage guides