Descriptions:
Web Dev Cody reviews Anthropic’s newly released Claude Opus 4.7, walking through its most significant changes compared to Opus 4.6 from the perspective of a developer actively using it in Cursor and Claude Code. The headline improvement is a major shift in instruction-following behavior: Opus 4.7 interprets prompts far more literally than prior versions, which means agent harnesses and prompts tuned for earlier Claude models may produce unexpected results and will need to be rewritten. The presenter describes firsthand experience repeatedly reprompting Opus 4.6 to actually complete all steps of a plan — a problem Opus 4.7 is specifically designed to fix.
Other notable changes include expanded multimodal vision support (images up to 2,576 pixels on the long edge, more than three times the previous limit), improved file-system-based memory for multi-session agentic work, and a new “extra high” effort level sitting between the existing “high” and “max” settings for finer latency-versus-reasoning control. On safety metrics, the model’s misaligned behavior score dropped from 2.75 to approximately 2.4, and coding benchmark performance improved from 77.8 to 80.5.
The video is practical in tone, aimed at developers already running Claude Code or Cursor who want to know whether upgrading warrants immediate attention — and specifically what the stricter instruction-following means for existing workflows. The presenter flags prompt injection resistance improvements as worth noting for anyone feeding external context from MCP servers or web sources into their agent pipelines.
📺 Source: Web Dev Cody · Published April 16, 2026
🏷️ Format: Review
