Introducing Claude Mythos PreviewAnthropicReleased March 15, 2026

Claude Mythos Preview

A preview of Anthropic’s next-generation autonomous reasoning system with 1M context, deployed for frontier safety evaluation.

reasoning-llmResearch Only$10.00 / 1M tok (Mythos Preview)Context: 1M (1,000,000 tokens)Arena ELO: 1510

Technical Specifications

Architecture Type
Autonomous Frontier Research Architecture
Total Parameters
Next-Gen Research Scale
Context Window
1M (1,000,000 tokens)
Max Output Tokens
128K (128,000 tokens)
Knowledge Cutoff
January 2026
Supported Modalities
text, code, vision, reasoning
License & Access
Anthropic Research Preview Terms

Benchmark Evaluations

Mmlu Pro
90.4
Gpqa Diamond
85
Swe Bench Verified
82.5
Chatbot Arena ELO
1510

Deep Architectural Overview

Claude Mythos Preview tested full-horizon autonomy and self-directed scientific experimentation under rigorous Responsible Scaling Policy (RSP) Tier 3 safety protocols.

Strengths & Considerations

Core Strengths
  • 1,000,000 token context window
  • 128k output buffer for comprehensive artifact generation
  • Advanced multi-day research synthesis
Known Limitations
  • Restricted preview access under Anthropic safety agreements

Token & API Pricing

Input Tokens (1M)$10.00
Output Tokens (1M)$50.00
Cached Input (1M)$1.0000
Pricing is verified directly against Anthropic's developer documentation and API rate sheets.

Similar & Alternative Models

Explore other frontier models from Anthropic and comparable reasoning engines.

Browse all models
Anthropic

Anthropic Claude model for long-running agentic coding and knowledge work.

1M ctxUSD 4 / 1M input tok; USD 20 / 1M output tok
Anthropic

The ultimate model for demanding reasoning and long-horizon agentic work, featuring 0.025x cache reads ($0.25/MTok).

1M ctx$10.00 / 1M tok (0.025x Cache Reads)
Anthropic

The most capable autonomous intelligence system engineered by Anthropic, restricted to Project Glasswing participants.

1M ctx$10.00 / 1M tok (Glasswing Exclusive)
Anthropic

The premier enterprise flagship model for multi-hour autonomous coding, complex system engineering, and vision-heavy agent tasks.

1M ctx$5.00 / 1M tok (Opus 5 Flagship)