Source: anthropic.com/news/claude-fable-5-mythos-5

Table of Contents

  1. Why it exists

  2. What it can actually do

  3. Before you start

  4. How to access it

  5. Setup step by step

  6. Prompts to use it

  7. Models and routing reference

  8. Daily workflow

  9. Tips

  10. Troubleshooting

  11. Quick reference

🧠 Why it exists

Anthropic kept hitting a wall. Their most capable models could do extraordinary things, but the same capabilities that help a cyberdefender secure software could help an attacker break it. So for months the strongest tier, called Mythos class, stayed locked to a small group under Project Glasswing.

Fable 5 is the answer to that wall. It is the same underlying model as Mythos 5, but wrapped in safeguards that make it safe enough to hand to everyone. When you ask something risky around cybersecurity, biology, chemistry, or model distillation, the request quietly falls back to Claude Opus 4.8 instead. Everything else runs at full Mythos class power.

The result: the most capable model Anthropic has ever released publicly, with a safety net built in.

⚡ What it can actually do

You say

What happens

"Migrate this entire codebase to the new framework"

Handles codebase-wide migrations. Stripe ran one on 50M lines of Ruby in a day that would have taken a team over two months by hand.

"Rebuild this app from these screenshots"

Reconstructs working source code from images alone, no other input.

"Analyze this financial document and pull the key figures"

Tops Hebbia's senior-level finance reasoning benchmark, strong on charts, tables, and document reasoning.

"Play this game and figure it out"

Beat Pokemon FireRed using vision only, with no maps or game-state helpers.

"Design a 3D-printable part"

Built a complete 3D-printable CAD model inside a browser-based editor it also created.

"Work on this long, multi-step task"

Stays focused across millions of tokens and improves its work using its own notes.

❝

⚠️ Many of the headline results above come from Anthropic's own early-access customers and internal tests. Treat them as vendor-reported until independently verified.

✅ Before you start

A quick checklist before you dive in:

  • A Claude account at claude.ai

  • Know which plan you are on (Free, Pro, Max, Team, or Enterprise)

  • For developers: access to the Claude API via the Claude Platform

  • Awareness that some requests will route to Opus 4.8 instead (this is normal, see routing reference)

🚀 How to access it

Fable 5 is not a tool you install. It is a model you select inside Claude or call through the API. There is no CLI, no package, no terminal command.

In the Claude app or web:

  1. Open claude.ai or the desktop or mobile app

  2. Start or open a chat

  3. Open the model selector

  4. Choose Claude Fable 5

Via the Claude API:

  1. Go to the Claude Platform console

  2. Use the model string claude-fable-5 in your API request

  3. Pricing is $10 per million input tokens and $50 per million output tokens

❝

⚠️ Availability on subscription plans is staged. From launch through June 22, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost. On June 23, using it on those plans requires usage credits. Anthropic says it intends to restore it as a standard plan feature once capacity allows. Verify current status before relying on this.

🛠️ Setup step by step

  1. Sign in to your Claude account.

  2. Open a new conversation.

  3. Click the model name at the top of the chat to open the model selector.

  4. Select Claude Fable 5 from the list.

  5. Confirm the selected model shows Fable 5 before sending important prompts.

  6. Start prompting. For long or complex work, give it the full task up front rather than feeding it piecemeal.

  7. If a response says it was handled by Opus 4.8, that means your request hit a safeguard and was routed. Rephrase if it was a false positive, or continue with the Opus answer.

💬 Prompts to use it

Grouped by use case. Copy, paste, adapt.

Software engineering

Here is my codebase. Migrate every file from [old framework] to [new framework].
Preserve behavior, update tests, and flag anything you could not migrate cleanly.
Here are screenshots of an app. Rebuild the front-end source code to match,
component by component, and tell me where you had to guess.

Knowledge work and finance

Here is a [financial report / contract / dataset]. Pull the key figures into a table,
flag anything unusual, and give me a one-paragraph summary of what matters.

Long-horizon tasks

This is a multi-day project. Keep notes as you go, work through it step by step,
and check your own output before telling me it is done.

Vision

Read the exact numbers off this chart and rebuild the underlying data as a table.

📊 Models and routing reference

Model

What it is

Who gets it

Claude Fable 5

Mythos-class model with safeguards on. Public release.

Everyone, today

Claude Mythos 5

Same model, cyber safeguards lifted.

Project Glasswing partners and US government

Claude Opus 4.8

Next-most-capable model. Handles Fable's flagged fallbacks.

Everyone

When Fable 5 routes to Opus 4.8 instead:

Trigger area

Why

Cybersecurity

Prevents uplift for cyberattacks

Biology and chemistry

Prevents uplift for dangerous bio research

Distillation

Prevents extracting Fable's abilities to train competing models

❝

⚠️ Anthropic states the fallback triggers in on average less than 5% of sessions, and that safeguards are deliberately tuned conservatively, so some harmless requests will get caught. This number may change as they refine the classifiers.

📅 Daily workflow

A real person using Fable 5 day to day:

  1. Morning: Open Claude, confirm Fable 5 is selected. Hand it the biggest, gnarliest task first, the one you would normally break into pieces. Fable's lead grows with task length.

  2. Midday: For analysis-heavy work, drop in your documents or screenshots and let it extract and summarize rather than reading manually.

  3. When it falls back: If you get an Opus 4.8 response on something harmless, rephrase once. If it keeps routing, the topic is inside a safeguard zone, move on.

  4. For dev work: Call claude-fable-5 via the API for automated or repeatable jobs. Watch your token spend at $10 in / $50 out per million.

  5. Watch the calendar: If you are on a subscription plan, note the June 22 cutoff for included access.

💡 Tips

  • Give it long tasks on purpose. The longer and more complex the job, the bigger Fable's advantage over other models. Short trivia questions waste it.

  • Let it keep its own notes on multi-step work. Persistent memory measurably improved its performance on long tasks in Anthropic's testing.

  • A fallback to Opus 4.8 is not a refusal. It is a different, still highly capable model answering. You lose Fable's edge, not your answer.

  • Feed it screenshots directly for UI or chart work. Vision is one of its standout strengths.

  • If you are cost-sensitive, batch your real work into the included window before June 23.

🔧 Troubleshooting

What you see

What it means

Fix

"This response was handled by Claude Opus 4.8"

Your request hit a safeguard classifier (cyber, bio/chem, or distillation)

Rephrase if it was a false positive. If it keeps routing, the topic is restricted.

Fable 5 not in your model list

Plan rollout is staged, or you are on a plan without current access

Check your plan. Subscription access is staged and changes around June 22 to 23.

Higher-than-expected API bill

Fable is priced at $10 in / $50 out per million tokens

Monitor token usage. Use Opus or Sonnet for tasks that do not need Fable's edge.

Harmless request keeps getting blocked

Safeguards are tuned conservatively and catch some benign queries

Rephrase, or run the task on Opus 4.8 directly.

📌 Quick reference

Model name (UI):     Claude Fable 5
Model string (API):  claude-fable-5
Tier:                Mythos class (above Opus)
Twin model:          Claude Mythos 5 (restricted, safeguards lifted)
Fallback model:      Claude Opus 4.8
Price:               $10 / million input tokens, $50 / million output tokens
Fallback areas:      cybersecurity, biology + chemistry, distillation
Fallback rate:       under 5% of sessions (vendor-reported)
Free window:         Pro / Max / Team / seat Enterprise, through June 22, 2026
Launched:            June 9, 2026
Access:              claude.ai or platform.claude.com

By The AI Leverage - Learn and master AI daily

Keep Reading