Claude Opus 4.7 Model Card

Article

• 272-page safety/system card for Claude Opus 4.7.
• Heavy focus on Mythos benchmark (Anthropic’s internal red-team eval).
• Documents model behavior, refusals, and safety mitigations in detail.
• Standard Anthropic responsible scaling policy disclosure.

Discussion

• 272 pages called out as redefining what a “model card” means.
• Mythos benchmark mentioned 331 times — reads as an ad for the eval framework.
• Haiku’s lack of update sparks theory: small models are cannibalizing high-end demand.
• Raw stats shared: ‘opus’ appears 809×, ‘mythos’ 331× in the document.

Discuss on HN