Mythos

TL;DR

The Mythos-class context window is 1 million input tokens with 128K maximum output tokens, confirmed for the generally available Claude Fable 5 and Claude Mythos 5, both released by 📝Anthropic on June 9, 2026. 📝Claude Mythos Preview — the restricted research model behind Project Glasswing — shares the same model family, but Anthropic has not published its exact context window.

Overview

A 1M-token context window is roughly 750,000 words, or about 3.4 million Unicode characters — enough to ingest an entire large codebase in a single request. That scale matters most for the Mythos-class defensive-cybersecurity use case, where understanding a project's full attack surface requires holistic comprehension rather than file-by-file analysis.

Confirmed Specifications

Confirmed by Anthropic's public communications as of June 9, 2026.

  • Context window: 1,000,000 input tokens (Fable 5 / Mythos 5)
  • Max output: 128,000 tokens
  • Fable 5 model ID: claude-fable-5
  • Pricing (Fable 5 / Mythos 5): $10 / input MTok, $50 / output MTok
  • Released: June 9, 2026
  • Fable 5 vs Mythos 5: same underlying model. Fable 5 routes high-risk queries (cybersecurity, biology, distillation) to Claude Opus 4.8; Mythos 5 lifts those safeguards in some areas for trusted cyber-defense partners.

Source: Anthropic — Claude Fable 5 and Claude Mythos 5.

Project Glasswing and the Preview

Claude Mythos Preview is the unreleased frontier model Anthropic considers too capable for public release. It powers Project Glasswing, Anthropic's effort to secure critical software. Confirmed:

  • 12 launch partners: Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks — plus 40+ additional critical-infrastructure organizations.
  • $100M in Mythos Preview usage credits, plus $4M in open-source security donations ($2.5M to Alpha-Omega/OpenSSF via the Linux Foundation, $1.5M to the Apache Software Foundation).

Source: Anthropic — Project Glasswing.

Cybersecurity Results (Confirmed)

Holding an entire project's source alongside its reasoning chain, tool outputs, and reproduction steps in one agentic session, Mythos Preview produced:

  • 181 successful autonomous exploits against Firefox's JavaScript engine, versus 2 for Claude Opus 4.6.
  • 595 tier 1–2 crashes in OSS-Fuzz evaluation, with tier 5 full control-flow hijacks on 10 separate fully patched targets.
  • 89% severity agreement with expert reviewers across 198 manually reviewed vulnerability reports.
  • Thousands of zero-day vulnerabilities across major operating systems and browsers, some that "survived decades of human review."

Source: Anthropic — Claude Mythos Preview.

Reported, Not Confirmed

Rumor or third-party estimates as of today — not confirmed by Anthropic:

  • Mythos Preview pricing of $25 / $125 per MTok. Anthropic states only that Mythos 5 costs "less than half" of Mythos Preview, implying Preview exceeded $20 / $100. [reported]
  • Mythos Preview model IDs, knowledge cutoff, and an April 7, 2026 preview launch date — not published by Anthropic. [reported]
  • Specific benchmark figures (93.9% SWE-bench Verified, 77.8% SWE-bench Pro, 82.0% Terminal-Bench 2.0) — circulated via leaks, not in Anthropic's released materials. [rumor]
  • ~10 trillion parameters, Mixture-of-Experts architecture, $5–15B training cost — from leaked internal documents and independent estimates, not Anthropic disclosures. [rumor]

Rumor sources: AI & Analytics Diaries, The Decoder.

FAQ

How big is the Claude Mythos context window?

The generally available Mythos-class models — Claude Fable 5 and Claude Mythos 5 — have a 1M-token input window with 128K max output, confirmed by Anthropic. Mythos Preview's exact window has not been published.

What's the pricing?

Fable 5 and Mythos 5 are $10 per million input tokens and $50 per million output tokens. Mythos Preview pricing was higher and is not officially confirmed.

How does the context window help with cybersecurity?

The 1M window lets the model hold an entire project's source, its reasoning chain, tool outputs, and reproduction steps in one agentic session — enabling whole-program vulnerability research rather than file-by-file analysis.

Related

Subjective

The headline isn't the number — it's that a 1M-token window plus Mythos-class reasoning makes whole-codebase work practical in a single pass. The same property that makes it a formidable vulnerability researcher makes it a formidable refactoring and audit partner. That's why Anthropic shipped it locked down: Fable 5 routes the dangerous questions to Opus 4.8, while Mythos 5 keeps the capability available to trusted defenders.

Contexts

Created with 💜 by One Inc | Copyright 2026