[KG] Claude Mythos Preview — moat

{
  "entity_slug": "claude-mythos-preview",
  "entity_name": "Claude Mythos Preview",
  "entity_type": "ai_model",
  "title": "Anthropic's Mythos Preview: Browsing beast, but can it escape GPT-3.5's shadow?",
  "narrative": "Anthropic's Claude Mythos Preview is a research model that excels at deep browsing (86.9% on BrowseComp) and OSWorld-Verified (79.6%). Developed by Anthropic and endorsed by the UK AI Safety Institute, it's already in use by Mozilla and XBOW. But the graph reveals a curious competitive landscape: Mythos is explicitly pitted against both GPT-3.5 and its own sibling, Claude Opus 4.6. While GPT-3.5 is a legacy target, the internal rivalry with Opus 4.6 suggests Anthropic is testing a cheaper or more specialized agent. The model's dependencies on Firefox and Windows hint at a consumer-facing browsing agent. Recent coverage shows Mythos can build working exploits in hours—a capability that likely triggered a 3.5x spike in CVEs after launch. This isn't just a benchmark play; it's a dual-use tool with regulatory attention.",
  "key_points": [
    "Anthropic developed Claude Mythos Preview; endorsed by UK AI Safety Institute.",
    "Competes with GPT-3.5 and Claude Opus 4.6—internal and external rivalry.",
    "Uses Firefox, Windows, and METR; adopted by Mozilla and XBOW.",
    "Strong on BrowseComp (86.9%) and OSWorld-Verified (79.6%).",
    "Linked to a 3.5x CVE spike post-launch; builds exploits in hours."
  ],
  "angle": "moat",
  "neighborhood_size": 12,
  "generated_at": "2026-07-28T09:01:32.607960+00:00"
}