Too dangerous or just too expensive? The real reason Anthropic is hiding Mythos

· ai cloud · Source ↗

TLDR

  • Anthropic’s Claude Mythos Preview is restricted to ~40 vetted orgs via Project Glasswing, with evidence pointing to both genuine security concerns and acute compute constraints.

Key Takeaways

  • Mythos autonomously found real zero-day vulnerabilities in testing, not just reproduced known ones – Anthropic’s Frontier Red Team documented this as the threshold justifying restricted release.
  • Pricing is $25/$125 per million input/output tokens with a 1M token context window and 128K max output; Anthropic committed $100M in compute credits to Glasswing participants.
  • A leaked internal draft described Mythos as “expensive to run” alongside “not yet ready for general release” – the only document pairing cost and access restriction directly.
  • Anthropic’s compute situation in April 2026 was strained: CoreWeave gap-fill deal, Google/Broadcom TPU capacity not arriving until 2027, and ~$500M chip design initiative all announced within days of the Glasswing launch.
  • Restricting agentic harnesses simultaneously, citing “outsized strain on systems,” reinforces that capacity limits were a live operational concern during the same window.

Hacker News Comment Review

  • Commenters were split on whether the zero-day capability claims are genuinely novel or incremental over existing frontier models like ChatGPT Pro, with skeptics arguing Anthropic inflates vulnerability counts to create marketing hype.
  • The “too dangerous” framing is widely read as convenient for a pending IPO, with multiple commenters suggesting the narrative positions Anthropic favorably regardless of whether security or cost is the real driver.
  • One concise synthesis from the thread: both factors are probably true simultaneously, since the KV cache scaling that yields the emergent capability also drives the extreme serving cost.

Notable Comments

  • @wood_spirit: argues that if Mythos were truly transformative, Anthropic could charge enough to make it economical – failure to do so suggests it is only incrementally better than Opus 4.7.
  • @waynecochran: “The KV cache scaling yields both the emergent power and requires the enormous capacity” – frames security and cost as structurally linked, not competing explanations.

Original | Discuss on HN