How to Use Claude Fable 5 and Mythos 5: Practical Guide

Anthropic released Claude Fable 5 and Claude Mythos 5 on June 9, 2026. Both sit in a new Mythos-class tier above Opus. They share the same underlying model, a 1M token context window, and the same API pricing. The difference is access and safeguards.

If you searched how to use Claude Fable 5, you probably want to know when it beats Opus 4.8, how to prompt it, and what to expect in the Claude app versus the API. This guide covers that.

Fable 5 vs Mythos 5: what you can actually use

Claude Fable 5 is generally available. You can use it in the Claude app (on supported plans), via the Claude API as claude-fable-5, and on AWS Bedrock, Vertex AI, and Microsoft Foundry.

Claude Mythos 5 is the same model with some safety classifiers lifted. It is not public. Access is limited to approved partners (for example Project Glasswing for cybersecurity work). Most readers should plan around Fable 5, not Mythos 5.

If you do not have Mythos access, Fable 5 is the right model for hard reasoning and long agent runs.

When to use Fable 5 instead of Opus 4.8

Use Claude Fable 5 when:

• The task runs for hours or days (research, multi-file refactors, long reports)
• You need a 1M token context window for large codebases or document sets
• You want the model to plan, execute, and check its own work
• You are building agentic workflows that loop until a goal is met

Stick with Opus 4.8 or Sonnet when:

• You need fast chat replies
• Cost per request matters more than maximum depth
• The task is a single-turn question or a short code fix

Fable 5 costs more per token than Opus ($10 per million input, $50 per million output). Use it for work where quality and persistence justify the premium.

How to turn on Fable 5 in Claude

In the Claude web or desktop app:

1. Open a new chat
2. Click the model picker at the top
3. Select Claude Fable 5

On some subscription plans, Fable 5 may be included for a limited window and then move to usage-based billing. Check your plan details in Settings if the model disappears from the picker.

For team use, set a default model policy so everyone knows when Fable 5 is allowed (for example architecture reviews yes, quick Slack summaries no).

Prompting Fable 5 for long-horizon work

Fable 5 is built for tasks that do not finish in one reply. Prompt it like a senior colleague with a brief, not like a search box.

Start with outcome and constraints

Goal: Refactor our billing module to support multi-currency invoices.
Constraints: No schema changes to the payments table. Keep backward compatibility.
Deliverables: PR-ready code, test plan, migration notes.
Stop when: All existing tests pass and you have listed open risks.

Give it permission to explore

Fable 5 works better when you allow file reads, test runs, and iteration. In Claude Code or API agent setups, enable tools instead of pasting entire repos manually.

Ask for self-verification

Add lines like:

• "Run tests before you say you are done"
• "List assumptions you made"
• "If blocked, propose two options with trade-offs"

Chunk enormous inputs

Even with 1M context, quality drops if you dump noise. Prefer: repo map, key files, error logs, and a clear question over raw data dumps.

Using Fable 5 in the API

Model ID: claude-fable-5

Basic pattern (pseudo-code):

response = client.messages.create(
  model="claude-fable-5",
  max_tokens=16000,
  messages=[{"role": "user", "content": "Your task brief here"}]
)

Plan for safety classifier fallbacks. Fable 5 can route sensitive topics (cybersecurity, biology, chemistry, distillation) to Claude Opus 4.8 instead of answering directly. In production:

• Detect when the response model differs from the request
• Log refusals and fallbacks for support teams
• Retry with a narrower, compliant prompt if needed
• Do not assume every API call used Fable 5 end to end

Use prompt caching for repeated system prompts and large static context. Anthropic offers a large discount on cached input tokens, which matters at Fable 5 pricing.

Set max output thoughtfully. Fable 5 supports very large outputs (up to 128k tokens per request). Long outputs cost more. Cap output unless you truly need a full document.

Mythos 5: who it is for

Mythos 5 is for vetted partners doing authorized cybersecurity or research work under Anthropic's trusted access programs. It is not a consumer upgrade.

If you are a normal product team, startup, or enterprise app builder, Fable 5 is your Anthropic frontier model. Do not wait for Mythos 5 unless your legal and security teams are already in an approved program.

Cost control checklist

• Use Fable 5 only on tasks you would give a senior engineer, not for every chat
• Summarize long threads before starting a new Fable session
• Cache stable instructions and docs in API integrations
• Monitor token usage per feature in your app dashboard
• Fall back to Sonnet or Opus for classification, tagging, and simple extraction

Fable 5 vs ChatGPT for the same task

Rough rule of thumb in June 2026:

See our ChatGPT 5.5 workflow guide for the OpenAI side.

Common mistakes

• Using Fable 5 for one-line questions (waste of money)
• Ignoring fallback to Opus 4.8 in API error handling
• Pasting entire repos without structure
• No definition of done, so the model keeps going past useful output
• Letting every employee use frontier models without a usage policy

Questions people ask

• Q: Is Fable 5 the same as Mythos 5? A: Same base model. Fable 5 has safety classifiers for general use. Mythos 5 has some limits removed and is restricted to approved access programs.
• Q: Why did my Fable 5 answer feel like Opus? A: Safety routing may have sent your request to Opus 4.8. Check response metadata in the API or try rephrasing within policy.
• Q: Do I need Fable 5 for RAG chatbots? A: Usually no. Sonnet or Opus plus good retrieval is enough. Fable 5 helps for complex agent loops and very large contexts.
• Q: Can I fine-tune Fable 5? A: Follow Anthropic's current API docs for your account. Most teams get better ROI from prompting, tools, and evals first.

Related reading

• Agentic AI Engineer career guide
• Prompt Engineer career guide
• Forward Deployed Engineer (FDE) career guide
• AI Engineering staffing hub
• How to use ChatGPT 5.5 workflows

Building production AI systems?

Frontier models are only part of the stack. You still need engineers who can wire models into your product with auth, evals, and observability. AllDomainSoft places AI engineers, FDEs, and agent builders in our Gurgaon office for US, UK, and Australian teams.

Request AI engineering profiles.