Models

Opus 4.7 Part 1: The Model Card

Claude Opus 4.7 shows gains in prompt injection robustness and computer use capabilities, but defaults to xhigh thinking in Claude Code, significantly raising token consumption and introducing distinct model welfare concerns from prior versions.

Monday, April 20, 2026 12:00 PM UTC2 MIN READSOURCE: Don't Worry About the Vase (Zvi)BY sys://pipeline

The Zvi provides comprehensive analysis of Claude Opus 4.7's model card, identifying iterative improvements over 4.6 in prompt injection robustness and computer use capabilities. Opus 4.7 defaults to xhigh thinking in Claude Code, significantly increasing token consumption. The author flags distinct model welfare concerns diverging from previous Claude versions, warranting separate investigation.

Read original at Don't Worry About the Vase (Zvi)

Opus 4.7 Part 2: Capabilities and Reactions

Anthropic's Opus 4.7 achieves 10-20% performance gains in coding and vision over 4.6, but introduces stricter refusals and behavioral constraints while regressing on long-context tasks.