Claude Fable 5: The First Public Mythos-Class Model — Developer Guide

Anthropic shipped Claude Fable 5 on June 9, 2026, and it changes the math on what you can expect from a generally available model. Until now, Mythos-class intelligence was locked behind a trust-tier waitlist. Now it's on the open API — with a few guardrails attached.

If you're deciding whether to migrate from Opus 4.8, setting up your first Fable call, or just trying to understand where this fits in the Claude family, this guide covers everything a developer needs to know in the first week after launch.

What Is Claude Fable 5 (And How Is It Different From Mythos?)

Claude Fable 5 is the publicly released variant of Claude Mythos 5. Both share the same underlying frontier model weights, but Fable ships with classifier-based safeguards that block responses in four high-risk domains: cybersecurity, biology, chemistry, and distillation. When a request triggers one of those classifiers, the API falls back to Claude Opus 4.8 for a safe answer — and returns stop_reason: "refusal" in the HTTP 200 response body so your application can detect the switch.

In 95% of sessions, Fable 5 runs end-to-end on the full Mythos-class model. The fallback fires rarely in mainstream development and business use cases. If your workload intentionally lives in the high-risk zone (offensive security research, wet-lab guidance), you'll need the restricted Claude Mythos 5 access, not Fable.

Benchmarks: Claude Fable 5 vs Claude Opus 4.8

Here's what the numbers look like across the benchmarks Anthropic published at launch:

The headline finding: Fable 5's lead grows with task length and complexity. On short, well-scoped coding tasks, the gap versus Opus 4.8 is modest. On long-horizon agentic work — multi-step refactors, end-to-end feature implementation, complex RAG pipelines — the advantage becomes material.

That benchmark pattern has direct architectural implications: Fable 5 is the right default when you're handing the model an ambiguous, multi-step goal. Opus 4.8 remains the cost-efficient choice for high-volume, predictable, bounded tasks.

API Setup: Calling Claude Fable 5 in 5 Minutes

Model ID

claude-fable-5

On Amazon Bedrock, the regional form is:

us.anthropic.claude-fable-5

On Vertex AI, use the standard Anthropic endpoint with model name claude-fable-5.

Python (Anthropic SDK)

pythonimport anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-fable-5",
    max_tokens=4096,
    messages=[
        {
            "role": "user",
            "content": "Refactor this service to use dependency injection and add integration tests."
        }
    ]
)

print(response.content[0].text)

TypeScript / Node.js

typescriptimport Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();

const message = await client.messages.create({
  model: "claude-fable-5",
  max_tokens: 8192,
  messages: [
    {
      role: "user",
      content: "Analyze this codebase and propose a migration plan to microservices."
    }
  ]
});

console.log(message.content[0].text);

One critical difference from Opus 4.8: thinking mode

Fable 5 uses adaptive thinking only. You cannot set thinking: { type: "disabled" } — the API will return an error. Depth of reasoning is controlled through the effort parameter instead:

pythonresponse = client.messages.create(
    model="claude-fable-5",
    max_tokens=16000,
    effort="high",          # "low" | "medium" | "high" — controls reasoning depth
    messages=[...]
)

Use effort="low" to optimize for speed and cost on simpler queries, effort="high" when you need maximum reasoning depth on complex tasks.

Handling refusals gracefully

When a request hits the safeguard classifiers, the response looks like this:

json{
  "stop_reason": "refusal",
  "content": [
    {
      "type": "text",
      "text": "I can't help with that request. [Safe Opus 4.8 response follows]"
    }
  ]
}

Always check stop_reason in your application layer if you need to distinguish a full Fable response from a fallback-triggered Opus response.

Pricing: What Fable 5 Actually Costs

Fable 5 costs exactly 2× Opus 4.8 at every tier. For workloads generating 10M output tokens per month, the cost difference is $250,000/year. That's real money — so the routing question matters.

When to Use Fable 5 vs Opus 4.8

The right call depends on task type, not raw capability:

• Long-horizon agentic workflows (multi-step coding agents, autonomous refactoring)
• Complex multi-document reasoning (research synthesis, legal analysis across large corpora)
• Hard coding tasks where SWE-Bench-style difficulty is the norm
• Multimodal analysis requiring grounded, detailed visual understanding
• Any task where you're currently hitting the quality ceiling on Opus 4.8

• High-volume, bounded tasks (classification, summarization, extraction)
• Workloads with latency-sensitive SLAs (Fable's adaptive thinking adds latency)
• Applications where cost is the primary constraint
• High-risk domain queries that will always fall back anyway (cybersecurity, bio, chem)

The smart architecture for most production systems: route by task complexity at the orchestration layer. Use Fable 5 as your "hard jobs" model and Opus 4.8 as the default. You get near-frontier intelligence where it matters, without paying frontier prices across the board.

Context Window and Output Limits

Claude Fable 5 matches Claude Mythos Preview on context:

• Context window: 1,000,000 tokens (1M)
• Max output tokens: 128,000 tokens per request

These are the same limits as Claude Sonnet 4.6 and Opus 4.8. The difference is that Fable 5 uses that context more effectively — the model's ability to reason coherently over a full 1M context is measurably stronger than Opus 4.8 on long-form tasks.

Where Claude Fable 5 Is Available

As of June 9, 2026:

• Claude API (direct Anthropic) — generally available
• Claude Platform on AWS — generally available
• Amazon Bedrock — generally available in US regions
• Google Vertex AI — generally available
• Microsoft Foundry — generally available
• GitHub Copilot — generally available
• Claude claude.ai interface — available on Pro and Max plans (no extra credit cost through June 22)

This is a notably broad launch-day availability. Past flagship models like Opus 4.8 launched on direct API first, with cloud provider integrations following 2–4 weeks later.

What This Means for Claude Certified Architect Candidates

If you're studying for the Claude Certified Architect (CCA) exam, Fable 5 is almost certainly going to appear in exam questions within the next certification refresh cycle. Key concepts to understand:

Our CCA practice test bank already includes questions on model selection, context management, and agentic architecture — the core domains where Fable 5 will show up on exams. The free CCA study guide covers the model family in depth.

Key Takeaways

• Claude Fable 5 is the first public Mythos-class model, released June 9, 2026 on all major platforms
• It uses the same underlying weights as Claude Mythos 5, with classifier safeguards in four high-risk domains
• +11 points on SWE-Bench Pro, ~2× on FrontierCode vs Opus 4.8 — advantages are largest on long, complex tasks
• Pricing: $10/$50 per 1M tokens — 2× Opus 4.8, so routing by task complexity matters
• API model ID: claude-fable-5 — adaptive thinking only, use effort parameter to control depth
• Free for Pro/Max/Team/Enterprise through June 22; usage credits required after June 23
• Handle stop_reason: "refusal" in production code for the Opus 4.8 fallback path

Next Steps

Start with the two-week free window: swap claude-opus-4-8 for claude-fable-5 in your most complex agentic workflow, run it against your real workload, and measure the quality delta before June 23. That empirical data point — on your specific tasks — is worth more than any benchmark.

For developers and engineers building Claude expertise for interviews and certifications, the CCA remains the benchmark certification for Anthropic ecosystem competency. Check out our free CCA practice questions to see where Claude Fable 5 fits in the exam blueprint.

• Claude Fable 5 and Claude Mythos 5 — Anthropic
• Introducing Claude Fable 5 — Claude API Docs
• Claude Fable 5 is generally available for GitHub Copilot — GitHub Changelog
• Anthropic's Claude Fable 5 — TechCrunch
• Claude Fable 5 vs Opus 4.8 Benchmarks — TrueFoundry