First Opus-class model with 1M token context window (beta), adaptive thinking with effort levels, and context compaction for sustained agentic tasks. Agent teams feature enables parallel subtask execution.