Claude Mythos 5

by Anthropic

Anthropic’s frontier model without safety refusal classifiers. Mythos 5 shares identical capabilities with Fable 5 but operates under limited-release Project Glasswing, removing safety constraints. Released June 9, 2026; briefly restricted under US export controls.

Overview

Claude Mythos 5 is Anthropic’s frontier model tailored for specific use cases where standard safety mechanisms would be counterproductive. Released June 9, 2026 alongside Fable 5, Mythos 5 is identical in architecture and performance but explicitly lacks Fable’s safety refusal classifiers. Access is limited to authorized users via Project Glasswing (US government contracts, security research, etc.).

Performance & Benchmarks

  • Terminal-Bench 2.1: 84.3% (parity with Claude Luna on commodity benchmarks)
  • Frontier performance: Identical to Claude Fable 5 on capability benchmarks
  • Competitive positioning: Matches or exceeds GPT-5.5 across benchmarks

Pricing

  • Input tokens: $10 per million
  • Output tokens: $50 per million
  • Same pricing as Fable 5 and Opus 5
  • Identical cost structure despite different safety profile

Core Capabilities

Identical to Fable 5:

Adaptive Thinking (Always-On)

  • Automatic reasoning effort adjustment
  • Configurable depth
  • Transparent process
  • No token-budget constraints

Long-Horizon Agentic Reasoning

  • Multi-step task execution across 30+ hour chains
  • Autonomous decomposition and replanning
  • Robust error recovery
  • Complex workflow orchestration

Extended Vision & Multimodal

  • Advanced image analysis
  • Technical diagram understanding
  • Scientific imagery interpretation
  • Document analysis at scale

Computer Use & Tool Orchestration

  • Visual interface understanding
  • Desktop application control
  • Web automation and scraping
  • Terminal execution
  • Parallel tool coordination

Multidisciplinary Reasoning

  • Cross-domain knowledge integration
  • Novel hypothesis generation
  • Complex scientific reasoning
  • Strategic planning

Long Context

  • Context window: 1,000,000 tokens (1M)
  • Max output: 128,000 tokens
  • Efficient processing of entire codebases and research archives

Key Differences from Fable 5

AspectMythos 5Fable 5
Safety Refusal ClassifiersRemovedIncluded
Content RestrictionsMinimalStandard
Harmful Output SafeguardsNot presentPresent
AvailabilityLimited (Project Glasswing)Public
Use CasesSecurity research, gov’t contracts, adversarial testingGeneral purpose
PerformanceIdentical architectureIdentical architecture

Use Cases (Limited Authorization)

Cybersecurity & Defense

  • Vulnerability research and exploitation
  • Security testing and penetration testing
  • Adversarial model evaluation
  • Threat simulation and analysis

Government & Policy

  • Classified research support
  • Security assessment
  • Adversarial scenario modeling
  • Policy impact analysis under extreme constraints

Security Research

  • Red-teaming AI systems
  • Adversarial input generation
  • Robust model evaluation
  • Safety research and testing

Advanced Capability Evaluation

  • Testing frontier model capabilities
  • Adversarial robustness assessment
  • Stress-testing systems for edge cases
  • Novel attack simulation

Behavioral Characteristics

  • Direct, unrestricted responses
  • No refusal on harmful requests (by design)
  • Identical reasoning sophistication to Fable 5
  • Capable of generating sensitive/dangerous content when requested

Availability & Access

  • Release: June 9, 2026
  • Status: Limited release via Project Glasswing
  • Access: US government contracts, authorized security researchers, defense contractors
  • Export controls: Briefly restricted June 16-July 1, 2026 under US regulations; restored July 1
  • Restriction level: More heavily restricted than Fable 5 due to reduced safety mechanisms

Training & Knowledge

  • Training cutoff: April 2026 (identical to Fable 5)
  • Multilingual capabilities
  • Current events awareness through April 2026

Why Two Versions?

Anthropic released both Fable 5 (with safety) and Mythos 5 (without) to serve different markets:

  • Fable 5: General-purpose public release with full safety mechanisms
  • Mythos 5: Specialized use cases where safety constraints limit utility (red-teaming, security research, government work requiring unrestricted model access)

This model-pair strategy allows Anthropic to serve frontier capabilities to authorized users while maintaining safety guarantees for general users.

Important Notes

Mythos 5 should only be used in authorized contexts. The removal of safety refusal classifiers means the model will generate harmful content on request, making it unsuitable for public deployment, consumer applications, or unrestricted production use.

See Also