# Introducing Claude Fable 5 and Mythos 5: The Adaptive Thinking Era

> Anthropic launches Claude Fable 5 and Mythos 5 with "Always On" adaptive thinking, a 1M token context, and new paradigms for long-horizon agentic workflows.

**Published:** 2026-06-09
**Category:** Model
**Automation impact score:** 98/100
**Substitution risk:** Critical

## Key takeaways

- Fable 5 is Anthropic's most capable widely released model for long-horizon agentic work.
- Adaptive Thinking is now strictly 'Always On', shifting reasoning control to an 'Effort' parameter.
- Mythos 5 offers the same capabilities but without safety classifiers, restricted to Project Glasswing.
- 1M token context remains standard, solidifying models as full-repository operators.

## What to do about it

- Transition prompt engineering strategies to leverage 'Effort' rather than explicit thinking triggers.
- Implement server-side or client-side SDK middleware to handle the new 'stop_reason: refusal' workflow.
- If you're in enterprise security, request access to Project Glasswing for Mythos 5 to test adversarial boundaries without classifier friction.

# Introducing Claude Fable 5 and Mythos 5

On June 9, 2026, Anthropic officially released **Claude Fable 5** and **Claude Mythos 5**. This launch represents a massive architectural shift for developers, focusing on long-horizon agentic capabilities and fundamentally changing how the model "thinks."

## The Twin Titans: Fable vs. Mythos

While both models share the same underlying intelligence and a massive **1M token context window**, they are built for different operational environments:

1. **Claude Fable 5:** The flagship, generally available model designed for the most demanding reasoning tasks. It includes active safety classifiers that can gracefully decline requests (returning a `stop_reason: "refusal"`).
2. **Claude Mythos 5:** An identical twin in raw capability, but stripped of safety classifiers. Mythos 5 is highly restricted, available only through the exclusive **Project Glasswing** program. It is engineered for specialized enterprise use cases, such as deep vulnerability research and red-teaming, where safety filters would hinder legitimate adversarial simulations.

## "Always-On" Adaptive Thinking

The most significant change in this generation is the deprecation of optional reasoning. **Adaptive Thinking is now always on.** 

You can no longer pass `thinking: {"type": "disabled"}`. Instead, developers must use the new **Effort** parameter to control the depth and cost of the model's reasoning. Furthermore, the raw "chain of thought" is completely hidden from the API response; developers can only choose to see a "summarized" version or omit it entirely.

## What This Means for Your Career

At Job Security Meter, we track the impact of AI capabilities on human roles. The introduction of Fable 5 pushes AI from a "conversational assistant" to a "long-horizon operator."

- **For Software Engineers:** Fable 5's native integration with code execution, task budgets, and memory tools means it can autonomously debug and refactor across massive codebases. You must pivot from "writing code" to "architecting systems and managing task budgets."
- **For Cybersecurity Professionals:** If your company is approved for Mythos 5, the model can simulate nation-state level attacks without being blocked by safety filters. Manual penetration testing is rapidly becoming obsolete.

## Migration and Pricing
Both models are priced at $10 per million input tokens and $50 per million output tokens. For integrations moving to Fable 5, Anthropic has introduced a robust "Fallback" mechanism. If Fable 5 refuses a prompt, developers can automatically retry on a different model (like Opus or Sonnet) and receive a prompt-cache refund, ensuring you don't pay twice for the same context.


---

**Source:** Job Security Meter — https://jobsecuritymeter.com/blog/introducing-claude-fable-5-and-mythos-5
**Human-readable version:** https://jobsecuritymeter.com/blog/introducing-claude-fable-5-and-mythos-5
**Last updated:** 2026-06-09
**Attribution:** Free to quote and cite with attribution to Job Security Meter.
**Full site specification for language models:** https://jobsecuritymeter.com/llms-full.txt