Claude Fable 5 and Mythos 5 Launch

·Toolin Editorial Team

Anthropic has launched its Mythos-tier flagship models: Fable 5 is open to all users and scores 80.3% on the software engineering benchmark SWE-bench Pro, with input priced at $10 per million tokens.

Claude Fable 5 and Mythos 5 Launch

Anthropic has released two Mythos-tier flagship models at once: Claude Fable 5 is open to all users, while Claude Mythos 5 is limited to trusted partners. Both are built on the same underlying model; the difference is that Fable 5 ships with built-in safety classifiers that automatically downgrade to Opus 4.8 when requests touch cybersecurity, biochemistry, or model distillation.

Official launch image of Fable 5 and Mythos 5

What Is Fable 5

Fable 5 is the strongest model Anthropic has ever publicly released. It surpasses its predecessor across software engineering, knowledge work, visual reasoning, long-context memory, and scientific research. The name Fable comes from the Latin fabula ("story"), related to Mythos ("myth") — the same underlying model, with the public edition called "fable" and the full edition called "myth."

One core advantage of Fable 5: the longer and more complex the task, the wider its lead over other models grows. It holds focus across million-token, long-running tasks and improves its output using its own notes.

Core Capabilities at a Glance

Software engineering: SWE-bench Pro 80.3%

  • SWE-bench Pro score of 80.3%, far ahead of GPT-5.5's 58.6% and Gemini 3.1 Pro's 54.2%
  • FrontierCode Diamond score of 29.3% (GPT-5.5 manages just 5.7%), and it hits that top score at medium reasoning effort
  • Stripe used Fable 5 to migrate an entire codebase of 50 million lines of Ruby in a single day (a job that would take humans more than two months)

Visual reasoning: beating Pokémon with pure vision

Previous Claude models needed elaborate tooling scaffolds to play Pokémon FireRed. Fable 5 beat the entire game from scratch on screenshots alone, without any map cheats or extra information.

On GDPpdf (a visual document reasoning benchmark), Fable 5 scored 29.8%, ahead of Opus 4.8 (22.5%) and GPT-5.5 (24.9%).

Long context and memory

Fable 5 holds focus across million-token, long-horizon tasks. In Slay the Spire playtests, once persistent file-based memory was attached, Fable 5's performance gain was 3x that of Opus 4.8. That means the model's ability to "learn from its own experience" has stepped up to a new level.

Long context and memory comparison

Scientific research (Mythos 5)

  • Protein design: produced 9 candidate designs for 14 disease targets, spanning immune, neurodegenerative, and muscle diseases
  • Genomics: autonomously collected data on 138 species and millions of cells, then trained a model 100x smaller that outperforms results published in Science

Pricing and Availability

ItemDetails
API nameclaude-fable-5
Input price$10 / million tokens
Output price$50 / million tokens
Free windowFrom now through June 22, free for Pro/Max/Team/Enterprise subscribers
From June 23Removed from subscription plans; requires usage credits

That's twice the price of Opus 4.8, but less than half of Mythos Preview.

Tip: Finish your heaviest tasks before June 22. After that, Fable 5 will be removed from subscription plans and available only through API credits. Anthropic says it will be readmitted to subscriptions when capacity allows.

Safety Mechanisms

Fable 5 sits behind a set of independent classifiers watching for three risk categories: cybersecurity, biochemistry, and model distillation. On a trigger, replies automatically downgrade to Opus 4.8. Official figures show over 95% of conversations never trip the downgrade.

Note: the classifiers are currently tuned conservatively, so legitimate code review and security-hardening requests can get caught in the crossfire. Anthropic says it will work on reducing the false-positive rate.

Money-Saving Tips for Fable 5

Fable 5's list price is double Opus's, but the real cost comes from how much it "thinks and how long it runs." Practical advice:

  1. Check the thinking level: switching to Fable from another model doesn't auto-reset the thinking setting — set it manually to what the task actually needs
  2. Compact sessions proactively: at each milestone in a large project, compact once (e.g. /compact) and don't let the session grow endlessly
  3. Scope tasks with a cheaper model first: before launching a large agentic run, use a cheaper model to pin down the task's scope and cut the number of rounds Fable spends feeling its way
  4. Default to Sonnet/Opus: everyday Q&A and simple code edits go to lower-tier models; switch to Fable only for genuinely complex multi-step tasks
  5. Watch your usage rhythm: under heavy use the 5-hour window can burn down in a few dozen minutes — schedule your most token-hungry tasks right after the window refreshes

Core principle: cheap models handle the everyday and the prep work; Fable handles the parts that truly demand a ceiling-level mind.

How to Get Started

  1. Head to claude.ai and pick Fable 5 in the model selector
  2. API users can call claude-fable-5 directly
  3. Claude Code users are supported as well

A good first move: test it on the most complex task you have at hand and get a direct feel for the gap versus Opus 4.8.

Related articles

wechat-cli: Manage Your WeChat Chat History from the Command Line
AI Products

wechat-cli: Manage Your WeChat Chat History from the Command Line

wechat-cli, an open-source WeChat command-line tool, offers 11 commands for operating on local WeChat data—export chat history, search messages, run statistical analysis—designed for AI Agent integration.

Toolin Editorial Team
Codepilot Beginner's Guide: The Open-Source Cross-Platform Agent Client
AI Tutorials

Codepilot Beginner's Guide: The Open-Source Cross-Platform Agent Client

The open-source Agent client Codepilot supports Claude, OpenAI, and Chinese models via Codeplan, with a full feature set including Skills, MCP, IM remote connections, and generative UI -- plus a detailed installation and configuration tutorial

Toolin Editorial Team
MemPalace: A Free Local AI Memory System That Is Blowing Up on GitHub
AI Products

MemPalace: A Free Local AI Memory System That Is Blowing Up on GitHub

The open-source AI memory system MemPalace uses the memory palace method, scores 96.6% on LongMemEval, runs locally, supports MCP integration, and costs only 10 USD a year

Toolin Editorial Team
DeepXiv: A CLI Tool That Lets AI Agents Directly Consume 200 Million Papers
AI Tutorials

DeepXiv: A CLI Tool That Lets AI Agents Directly Consume 200 Million Papers

DeepXiv is an open-source CLI tool that turns 200 million+ open papers into data interfaces Agents can call, supporting search, progressive reading, trending-topic tracking, and deep research.

Toolin Editorial Team
The Six Core Components of a Coding Agent, Fully Explained
AI Tutorials

The Six Core Components of a Coding Agent, Fully Explained

A breakdown of the core architecture of AI coding agents like Claude Code and Codex CLI, covering repository context, prompt caching, tool calling, context slimming, session memory, and subagent delegation.

Toolin Editorial Team
Graphify: Turn Any Folder into a Knowledge Graph with One Command
AI Products

Graphify: Turn Any Folder into a Knowledge Graph with One Command

A zero-configuration, all-modal, locally run open-source knowledge graph tool that cuts token consumption by 71.5x, needs no vector database, and installs with one pip command.

Toolin Editorial Team