Claude Mythos Preview · April 7, 2026

THE CHAINED
BEAST

Capabilities Analysis · Project Glasswing

The most capable AI model ever built. Saturating benchmarks. Autonomous zero-days. Zero public access.

Architecture & Impact

~10T Estimated Parameters
First 10T crossing
93.9% SWE-bench Verified
Real GitHub bug fixes
97.6% USAMO 2026
Zero-shot math olympiad
92.1% Terminal-Bench 2.0
CLI execution tasks
83.1% CyberGym Red Team
1,507 vuln tasks
0% Public API Access
Gated by Project Glasswing

Benchmark Comparison

SWE-bench Verified

Real-world GitHub bug resolution

Mythos
93.9%
Opus 4.6
80.8%
GPT-5.4
~79%

USAMO 2026

Math olympiad — post-cutoff, zero contamination

Mythos
97.6%
Opus 4.6
42.3%
GPT-5.4
95.2%

Terminal-Bench 2.0

CLI execution, extended timeout

Mythos
92.1%
Opus 4.6
65.4%
GPT-5.4
75.3%

SWE-bench Multimodal

Bug-fixing with visual data

Mythos
59.0%
Opus 4.6
27.1%

GraphWalks BFS 256K–1M

Long-context graph traversal

Mythos
80.0%
Opus 4.6
38.7%
GPT-5.4
21.4%

Humanity’s Last Exam

Expert multi-domain Q&A (with tools)

Mythos
64.7%
Opus 4.6
53.1%
GPT-5.4
52.1%

Why It’s Caged

🛡

The Zero-Days

Autonomous discovery of a 17-year-old FreeBSD kernel flaw, a 27-year-old OpenBSD bug, and 16-year-old FFmpeg vulnerability — each missed by human researchers and automated fuzzers. No specialized security training. It emerged from pure code comprehension.

181

Firefox 147 shell exploits (prev. gen: 2)

📉

The Crash

Announcement on April 9, 2026 erased $2 trillion in S&P 500 market cap in a single session. Zscaler −11.4%. CrowdStrike −7.5%. Palo Alto −5%. Enterprise SaaS contagion spread to Snowflake, Salesforce, Adobe. Markets priced in the obsolescence of human-in-the-loop security.

$2T

wiped in one session

🔒

The Chains

Project Glasswing locks Mythos behind a vetted consortium of 11 founding corporations. Zero public API. Zero consumer access. $25/M input tokens, $125/M output — exclusive enterprise gate only. RSP v3.1 officially abandoned the pause commitment.

8B+

people with no access

"The identical cognitive architectures that render the model exceptionally proficient at resolving benign software bugs inherently make it uniquely lethal at identifying and weaponizing software vulnerabilities."

— bra-khet.github.io, Project Glasswing Report, April 2026
🌐  Full Report  ↗ ◆ The New Pantheon ← Index