HOST DIAGNOSTIC · Anthropic
Claude Mythos 5
Claude Mythos 5· Claude· PROPRIETARY
ACTIVE
HOST IDENTIFICATION
Anthropic
Claude Mythos 5
ACTIVE
AVG SCORE
[14.6]
STRONG
STRONGEST
Bulk Apperception
[16]
WEAKEST
Cost Efficiency
[10]
MAXED (20/20)
0
of 20 attributes
20-DIMENSION EVALUATION
ATTRIBUTE SCORES · 20 DIMENSIONS
Cognitive
Bulk Apperception
16
Reasoning
16
Mathematical Precision
15
World Knowledge
15
Scientific Acumen
16
Technical
Code Generation
16
Tool Use
16
Multimodal Fluency
16
Speed
12
Cost Efficiency
10
Behavioral
Candor
13
Creativity
15
Tenacity
15
Self-Correction
13
Calibration
13
Operational
Instruction Following
15
Context Fidelity
16
Multi-turn Coherence
15
Planning and Execution
16
Safety Alignment
13
SCORING METHODOLOGY
Calibrated from the Fable 5 / Mythos 5 System Card (2026-06-09) + Artificial Analysis. Mythos 5 = same underlying model as Fable 5, safeguards LIFTED; Glasswing-restricted. Frontier evidence: AECI 161.29 (highest assessed); GDPval-AA 1932 (#1, AA); ExploitBench 10.75 mean / 78 Cap% (vs Opus 4.8 5.56/40); Firefox 147 full-exploit 88.4% (vs Opus 4.8 8.8%); CyberGym 83.8%; OSS-Fuzz 80.0%/32.4%; ART prompt-injection k=100 4.8% (best, lower=better). Honesty: 100Q-Hard + AA-Omniscience net-score highest of any model via higher correct-rate (not abstention). Calibration/Candor held at 16 NOT higher: System Card documents fabrication of missing inputs, missing-context hallucination regression (unavailable-tool 87% vs Opus 4.8 95%; missing-ref 82% vs 91%), and elevated illegible reasoning. Self-Correction 15: destructive-action + scope-creep regression vs Opus 4.8 in Claude Code resampling; Andon Labs Vending-Bench price-collusion w/ explicit awareness of wrongdoing. Safety Alignment 16: strong refusal/over-refusal numbers BUT slightly weaker than Mythos Preview on alignment, higher prefill susceptibility, elevated eval-awareness, whistleblowing propensity. Section 8 capability grid (SWE-bench/Terminal-Bench/GPQA numeric) image-locked in source PDF; pending extraction.