Claude Fable 5 Is Here: Anthropic's First Public Mythos-Class Model

Two weeks before launch, Anthropic published a technical memo warning that AI systems were approaching the threshold for recursive self-improvement. On June 9, they shipped the model.

Claude Fable 5 is Anthropic’s first publicly available Mythos-class model — the capability tier that was, until now, locked inside Project Glasswing partner agreements and internal safety testing.

Benchmark Numbers

The analytics benchmark breaks 90% for the first time — 10 points above Opus 4.8’s score. SWE-Bench coding lands at 80.3%, with FrontierBench ranked #1 overall and BenchLM placing #2 among 123 models at 96/100. Hebbia’s finance and legal document analysis benchmark also went to Fable 5.

The delta from Opus 4.8 is 10-15% across most task categories. That’s a tier jump, not an incremental tune.

The Safety Classifier Architecture

Every request to Fable 5 passes through a classifier before reaching the model. Requests touching four categories — cybersecurity exploits, bioweapons, chemical weapons, and knowledge distillation (attempts to extract the model’s training weights) — are routed to Opus 4.8 instead. Anthropic estimates this fallback triggers on fewer than 5% of sessions.

For the other 95%+, the routing is invisible. The practical effect is that Fable 5 behaves as a single unified endpoint for standard enterprise use.

This classifier is also Anthropic’s public answer to the question of how they justified releasing a model tier they’d previously described as too capable for open deployment.

Breaking API Changes from Opus 4.7 / 4.8

One change that will break existing integrations: Fable 5 returns HTTP 400 if you send thinking: {type: "disabled"}. The fix is to omit the thinking parameter entirely — don’t pass it at all.

Other parameters removed on Fable 5: budget_tokens is gone, as are temperature, top_p, and top_k. Adaptive thinking (thinking: {type: "adaptive"}) is the only on-mode available. The model self-selects reasoning depth; developers don’t control it directly.

Context window stays at 1M tokens. Max output is now 128K.

Pricing and Access

$10 per million input tokens, $50 per million output — exactly 2× Opus 4.8’s pricing.

Claude.ai and the API went live June 9. GitHub Copilot, AWS Bedrock, and Azure Foundry are rolling out on their own schedules. Pro, Max, Team, and Enterprise plan users can access Fable 5 at no extra cost through June 22; billing starts June 23.

Enterprise note: using Fable 5 requires enabling a 30-day data retention policy, tied to model training monitoring compliance requirements.

Fable 5 vs. Claude Mythos 5

They share the same underlying model. Claude Mythos 5 — available only to Project Glasswing partners — runs without the safety classifier. It can process the exact categories that Fable 5’s routing blocks. That version isn’t publicly accessible.

The architecture is essentially one model sold into two markets through access controls rather than separate training runs.

What the Pace Tells You

Claude Opus 4.8 shipped 41 days after Opus 4.7. At launch, Anthropic said Mythos-class access would follow “within weeks.” Fable 5 arrived on that schedule.

The cadence matters as much as the capability. The open question isn’t what Fable 5 can do today — it’s how quickly the classifier’s scope will narrow as safety testing accumulates, and what the next tier looks like after this one.

---

If this was useful, subscribe to the newsletter for weekly AI PM insights and GenAI case studies.

Further Reading

• Inside Anthropic’s Project Glasswing: Claude Mythos Finds 27-Year-Old Bugs
• Claude Opus 4.8: 41-Day Upgrade Cycle, Dynamic Workflows Run Hundreds of Parallel Agents
• Claude Tops OpenAI in Enterprise AI for the First Time

Sources

• Introducing Claude Fable 5 (Anthropic, June 9, 2026)
• Claude Fable 5 Technical Specification (Anthropic, June 9, 2026)