Claude Fable 5 and Claude Mythos 5: Anthropic's Mythos-Class Models Explained

Claude Fable 5 and Claude Mythos 5: What Anthropic Just Released

On June 9, 2026, Anthropic announced two of its most capable AI models to date: Claude Fable 5 and Claude Mythos 5. Both belong to a new tier Anthropic calls Mythos-class — a generation above Opus designed for demanding reasoning, long-horizon agentic work, and tasks that previous models could not sustain across hours or days.

Here is the short version: Fable 5 and Mythos 5 share the same underlying model. Fable 5 is the public, safeguarded version available to developers and enterprises everywhere. Mythos 5 is the same intelligence with safety restrictions lifted in approved domains — restricted to vetted partners through Project Glasswing and select research programs.

If you are a backend engineer wiring up agent workflows, a frontend developer evaluating AI copilots, or a platform team deciding which model to route production traffic to, this release changes the calculus. Fable 5 is not a marginal upgrade over Opus 4.8. It is a different class of model — with different pricing, different safety behavior, and different data retention rules you need to understand before shipping.

What Is a Mythos-Class Model?

Anthropic's model lineup has followed a clear ladder for years: Haiku for speed, Sonnet for balance, Opus for hard problems. Mythos-class sits above Opus — built for work that is not just difficult, but long.

Think of the difference this way. Opus 4.8 can solve a hard coding problem in a single session. Fable 5 can run autonomously across a multi-stage migration — reading thousands of files, updating dependencies, writing tests, verifying its own output, and continuing for extended periods without losing thread. That is the defining characteristic of Mythos-class models: sustained, self-directed execution on complex knowledge work.

Anthropic began previewing this capability in April 2026 with Claude Mythos Preview, limiting access to a small group of cybersecurity and infrastructure partners. Fable 5 represents the first time Mythos-level capability has been made safe enough — through classifiers, fallback routing, and data retention policies — for general commercial use.

What Mythos-Class Models Are Built For

• Large-scale code migrations across monorepos and legacy systems
• Multi-day agentic sessions with minimal human intervention
• Deep analytical research across long document sets
• Complex financial, legal, and scientific reasoning
• Vision-heavy tasks requiring sustained multimodal analysis
• Production-grade software engineering held to maintainability standards

Claude Fable 5 vs Claude Mythos 5: Same Brain, Different Guardrails

The relationship between Fable 5 and Mythos 5 is straightforward but easy to misunderstand. They are the same underlying model. The difference is entirely in the safety layer wrapped around Fable 5.

Anthropic describes Mythos 5 as having the strongest cybersecurity capabilities of any model in the world — which is precisely why general public access requires safeguards. Fable 5 is Anthropic's answer to the question: how do you bring Mythos-level capability to everyone without creating unacceptable misuse risk?

Key Capabilities: Long-Horizon Agentic Work and Coding

Fable 5 introduces what Anthropic calls its fifth model generation. The headline capabilities go beyond benchmark scores — they change what kinds of problems you can hand to an AI agent and trust the results.

Autonomous, Long-Running Tasks

Fable 5 can work autonomously for longer than any previous Claude model. It proactively self-updates skills based on learnings, develops its own evaluation harnesses, and verifies its work before delivery. For backend engineering teams, this means agent workflows that previously required constant human checkpointing can run with substantially more independence.

Software Engineering

Coding is the headline use case. Partners including Cursor, GitHub, and Cognition reported that Fable 5 handles complex, long-horizon coding tasks with a level of autonomy and reliability that exceeded previous benchmarks. It scores highest on CursorBench and leads on FrontierBench — Cognition's evaluation for frontier coding that tests generalization to unfamiliar tools.

Knowledge Work and Vision

Beyond code, Fable 5 targets professional tasks in finance, legal, marketing, sales, data analysis, and engineering. It is the first model to break 90% on Anthropic's core analytics benchmark for complex, long-running analytical tasks — a 10-point jump over Opus 4.8. Vision capabilities support document analysis, UI understanding, and multimodal workflows relevant to frontend engineering and design systems work.

Supported API Features at Launch

• Adaptive thinking (always on — cannot be disabled)
• Effort parameter for controlling reasoning depth
• Task budgets (beta)
• Memory tool
• Code execution
• Programmatic tool calling
• Tool result clearing through context editing (beta)
• Compaction
• Vision
• 1M token context window with up to 128k output tokens

Benchmark Performance: How Fable 5 Compares to Opus 4.8 and Competitors

Anthropic reports that Fable 5 is state-of-the-art on nearly all tested benchmarks, with the largest leads appearing on longer and more complex tasks. Third-party evaluations from early partners reinforce the gap.

Three numbers matter most for engineering teams. First, the +11-point jump on SWE-bench Pro over Opus 4.8 is large enough to change agent reliability on real coding tasks. Second, Fable 5 more than doubles Opus 4.8 on FrontierCode Diamond — a benchmark built around production-codebase standards for maintainable agentic coding. Third, Fable 5 reportedly achieves strong results even at medium reasoning effort, meaning you do not always need maximum token burn to get frontier-quality output.

As always, validate on your own workload. Public benchmarks are directional. Your codebase, your constraints, and your evaluation harness are what matter for production decisions.

Safety Classifiers, Refusals, and the Opus 4.8 Fallback System

This is the most important engineering detail in the Fable 5 release — and the one most integration guides skip. Fable 5 ships with safety classifiers: separate AI systems that run on every request to detect potential misuse, including jailbreak attempts.

Safeguarded Domains

Classifiers are tuned to block or redirect requests in high-risk areas:

• Cybersecurity: Exploitation, offensive cyber tasks, agentic hacking workflows
• Biology and chemistry: Content that could enable dangerous research
• Model distillation: Attempts to extract or replicate model capabilities

When classifiers trigger, Fable 5 does not return an error. It returns a normal response with stop_reason: "refusal". Your application must handle this — either by presenting the refusal to the user or by falling back to another model.

The Fallback Mechanism

Anthropic's recommended approach: configure automatic fallback to Claude Opus 4.8. Flagged queries get answered by Opus instead, and you are not charged Fable 5 pricing for rerouted requests.

Server-side fallback is available on the native Claude API and Claude Platform on AWS. Pass the server-side-fallback-2026-06-01 beta header with a fallbacks array. Amazon Bedrock, Vertex AI, and Microsoft Foundry require client-side fallback using SDK middleware.

Anthropic states the fallback triggers in under 5% of sessions during testing, and that safeguards are deliberately conservative — meaning benign technical work in adjacent domains sometimes triggers them. Expect false positives, especially early after launch, and plan your integration accordingly.

API Integration: Using claude-fable-5 in Node.js and TypeScript

For backend engineers building agent systems with Node.js or TypeScript, here is a production-ready pattern with server-side fallback configured:

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

const client = new Anthropic();

const response = await client.beta.messages.create({
  model: "claude-fable-5",
  max_tokens: 16384,
  betas: ["server-side-fallback-2026-06-01"],
  fallbacks: [{ model: "claude-opus-4-8" }],
  messages: [
    {
      role: "user",
      content: "Migrate this Express.js API to Fastify. "
        + "Preserve all route handlers and middleware behavior. "
        + "Write tests for each migrated endpoint.",
    },
  ],
  tools: [
    {
      name: "read_file",
      description: "Read a file from the project",
      input_schema: {
        type: "object",
        properties: {
          path: { type: "string" },
        },
        required: ["path"],
      },
    },
  ],
});

// Check which model actually served the response
const servedBy = response.model;
const wasRefusal = response.stop_reason === "refusal";

if (response.usage?.iterations) {
  for (const iteration of response.usage.iterations) {
    console.log(`Attempt: ${iteration.type} via ${iteration.model}`);
  }
}

For client-side fallback on platforms without server-side support — Bedrock, Vertex AI, Microsoft Foundry — use the SDK refusal-fallback middleware:

import Anthropic from "@anthropic-ai/sdk";
import { withRefusalFallback } from "@anthropic-ai/sdk/helpers";

const client = withRefusalFallback(
  new Anthropic(),
  { fallbacks: ["claude-opus-4-8"] }
);

const response = await client.beta.messages.create({
  model: "claude-fable-5",
  max_tokens: 8192,
  messages: [{ role: "user", content: userQuery }],
});

If you are building React frontends that call your own backend proxy, never expose API keys client-side. Route all Fable 5 requests through your Node.js or serverless API layer where you can enforce fallback logic, log which model served each response, and manage cost controls centrally.

Pricing, Context Window, and Data Retention Requirements

Pricing

Both Fable 5 and Mythos 5 cost $10 per million input tokens and $50 per million output tokens — roughly double Opus 4.8's pricing. That price alone acts as a natural throttle against sending every request to the frontier model. Prompt caching retains the existing 90% input token discount.

Context and Output Limits

• Context window: 1 million tokens by default
• Max output: 128,000 tokens per request
• Thinking: Adaptive thinking is always on and cannot be disabled via thinking: {"type": "disabled"}
• Effort parameter: Controls reasoning depth — use medium effort for many tasks to balance quality and cost

Mandatory 30-Day Data Retention

Fable 5 and Mythos 5 are designated covered models, which triggers mandatory 30-day data retention on all traffic — including third-party surfaces like Amazon Bedrock, GitHub Copilot, and Google Vertex AI. This overrides prior zero-retention agreements. There is no opt-out.

Anthropic states retained data is not used for model training, human access is logged, and data is deleted after 30 days in almost all cases. The stated purpose is defensive: catching novel attacks, multi-request abuse patterns, and new jailbreaks while reducing classifier false positives.

If your organization operates under strict data residency or zero-retention DPAs, review contracts and compliance requirements before routing production traffic to Fable 5. This is a vendor risk decision, not just a technical one.

Where to Access Claude Fable 5 and Mythos 5

Claude Fable 5 is generally available across Anthropic's distribution channels:

• Claude API: Direct access via claude-fable-5
• Amazon Bedrock: Unified AWS-managed access with Guardrails, Knowledge Bases, and regional data residency
• Claude Platform on AWS: Anthropic's native platform experience with unified AWS billing
• Google Vertex AI: Enterprise deployment through Google Cloud
• Microsoft Foundry: Access through Microsoft's AI platform
• Claude subscriptions: Included in Pro, Max, Team, and seat-based Enterprise plans through June 22, 2026 at no extra cost, then transitioning to usage credits
• Consumption-based Enterprise plan: Available for organizations with ambitious knowledge and coding workloads

Claude Mythos 5 is not generally available. Access is limited to approved Project Glasswing partners — cybersecurity organizations, infrastructure providers, and select biology researchers — with broader trusted-access programs planned.

Project Glasswing and Trusted Access for Claude Mythos 5

Project Glasswing is Anthropic's controlled deployment program for Mythos-class models without public safety restrictions. It launched with cybersecurity partners and expanded to hundreds of organizations across 15 countries, focusing on entities that manage critical infrastructure.

Mythos 5 serves as the upgrade path for organizations already on Claude Mythos Preview. Anthropic reports Mythos 5 is comparable to or somewhat stronger than Mythos Preview in most cases while costing substantially less than the preview pricing.

In consultation with the US government, Anthropic plans to steadily expand access through periodic additions to Glasswing and a broader trusted-access program that allows cybersecurity organizations to apply systematically. Biology researchers with approved access will receive Mythos 5 with biology and chemistry safeguards lifted.

For most developers and enterprises, Fable 5 is the correct model. Mythos 5 is relevant only if your organization is already approved for high-stakes security or research workloads where safeguard removal is both necessary and authorized.

When to Use Fable 5 vs Opus 4.8: A Smart Routing Strategy

At double the price of Opus 4.8, sending every request to Fable 5 is a costly mistake. The pattern most production teams adopt is complexity-based routing — a core system design principle applied to LLM infrastructure.

Routing decision tree:

Is the task long-horizon or multi-stage?
  YES → Is it business-critical?
    YES → Route to claude-fable-5
    NO  → Start with claude-opus-4-8, escalate on failure
  NO → Is latency-sensitive or high-volume?
    YES → Route to claude-sonnet-4-6 or claude-opus-4-8
    NO  → Route to claude-opus-4-8

Route to Fable 5 when:

• Large code migrations spanning thousands of files
• Multi-stage agent workflows requiring sustained autonomy
• Deep research across long document corpora
• Complex analytical tasks where judgment and nuance matter
• Tasks where Opus 4.8 has already failed or produced insufficient results

Keep on Opus 4.8 or smaller models when:

• Routine code generation, review, and documentation
• High-volume API endpoints with latency requirements
• Standard CRUD implementations and boilerplate
• Queries likely to trigger Fable 5 safety classifiers anyway

Build this routing into your API gateway or LLM proxy layer. Tools like model routers, task classifiers, or simple heuristics based on token count and task type can save significant cost without sacrificing quality where it matters.

What This Means for Backend Engineers, Frontend Teams, and DevOps

Backend Engineering

Fable 5 changes what you can delegate to agents. Multi-file refactors, API migrations, test suite generation across large codebases, and infrastructure-as-code updates become more viable as long-running autonomous tasks. Your job shifts toward defining scope, setting evaluation criteria, and reviewing output — the same architect-over-coder transition happening across the industry.

Frontend Engineering

For React and TypeScript teams, Fable 5's vision capabilities and long context window support UI audits, design system migrations, accessibility reviews across large component libraries, and end-to-end test generation. Pair it with your existing CI/CD pipeline by having agents verify their own changes against your test suite before opening pull requests.

DevOps and Cloud

On Amazon Bedrock and Claude Platform on AWS, Fable 5 integrates with existing AWS Guardrails and Knowledge Bases. DevOps teams should update monitoring to track Fable 5-specific metrics: refusal rates, fallback frequency, cost per task, and data retention compliance. Cloud cost management becomes more important when frontier models cost $50 per million output tokens.

System Design Considerations

Designing agent systems around Fable 5 requires thinking about failure modes that did not exist with previous models: classifier refusals mid-workflow, 30-day data retention implications for sensitive workloads, and cost runaway on long-horizon tasks. Build checkpointing, budget caps, and model fallback into your architecture from day one — not as an afterthought.

Key Takeaways

• Claude Fable 5 and Claude Mythos 5 are Anthropic's first Mythos-class models — a tier above Opus built for long-horizon agentic work.
• They share the same underlying model. Fable 5 adds safety classifiers for general use. Mythos 5 removes restrictions for approved partners.
• Fable 5 leads on agentic coding benchmarks — 80.3% on SWE-bench Pro, more than 2x Opus 4.8 on FrontierCode Diamond.
• Integrations must handle refusals and configure Opus 4.8 fallback — server-side on Claude API, client-side on Bedrock and Vertex.
• Pricing is $10/$50 per million input/output tokens, with mandatory 30-day data retention overriding zero-retention agreements.
• Smart routing — Fable 5 for hard long-horizon tasks, Opus 4.8 or Sonnet for everything else — is essential for cost control.
• Mythos 5 remains restricted to Project Glasswing and trusted-access programs for cybersecurity and approved research.

Frequently Asked Questions About Claude Fable 5 and Mythos 5

What is Claude Fable 5?

Claude Fable 5 is Anthropic's first generally available Mythos-class model, released June 9, 2026. It is designed for demanding reasoning, long-horizon agentic work, coding, knowledge work, and vision tasks. It includes safety classifiers that redirect high-risk queries to Claude Opus 4.8. The API model ID is claude-fable-5.

What is the difference between Claude Fable 5 and Claude Mythos 5?

They share the same underlying model and capabilities. Fable 5 wraps the model in safety classifiers for cyber, biology, chemistry, and distillation domains. Mythos 5 removes those restrictions for approved partners in Project Glasswing. Fable 5 is publicly available. Mythos 5 requires vetted access.

How much does Claude Fable 5 cost?

$10 per million input tokens and $50 per million output tokens — roughly double Opus 4.8. Prompt caching provides a 90% discount on cached input tokens. Requests rerouted to Opus 4.8 by safety classifiers are charged at Opus pricing, not Fable pricing.

Is Claude Fable 5 better than Opus 4.8?

Yes, for long-horizon and complex tasks. Fable 5 scores +11 points higher on SWE-bench Pro and more than doubles Opus 4.8 on FrontierCode Diamond. For routine, latency-sensitive, or high-volume work, Opus 4.8 at half the price is usually the better choice.

What is the context window for Claude Fable 5?

1 million tokens by default, with up to 128,000 output tokens per request. Adaptive thinking is always enabled and cannot be disabled.

What happens when Claude Fable 5 refuses a request?

Fable 5 returns a response with stop_reason: "refusal" rather than an error. Configure server-side or client-side fallback to Claude Opus 4.8 to automatically retry refused requests. You are not charged Fable pricing for rerouted requests.

Does Claude Fable 5 support zero data retention?

No. Fable 5 and Mythos 5 are covered models requiring mandatory 30-day data retention on all traffic, including third-party platforms. This overrides prior zero-retention agreements. Data is not used for training and is deleted after 30 days in most cases.

How do I get access to Claude Mythos 5?

Mythos 5 is not generally available. Access is limited to approved Project Glasswing partners for cybersecurity and infrastructure work, select biology researchers, and future trusted-access program applicants. Contact your Anthropic, AWS, or Google Cloud account team if your organization qualifies.

Where is Claude Fable 5 available?

Claude API, Amazon Bedrock, Claude Platform on AWS, Google Vertex AI, Microsoft Foundry, and Claude subscription plans. API access uses the model ID claude-fable-5.

When should I use Claude Fable 5 instead of other Claude models?

Use Fable 5 for ambitious, long-running, asynchronous tasks: large code migrations, multi-stage agent workflows, deep research, and complex analytical work. Use Opus 4.8 or Sonnet for routine coding, high-volume endpoints, and latency-sensitive applications. Route by task complexity, not by default.