2026-07-24 · Anthropic

v2.1.219

securityprotocolsagentsmodelscapital

read at source ↗ github.com

v2.1.219

Source: Anthropic Claude Code Date: 2026-07-24 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.219

Summary

Claude Code v2.1.219 makes Claude Opus 5 (1M context, fast mode at $10/$50 per Mtok) the default Opus model, and adds a sandbox.network.strictAllowlist setting that denies non-allowlisted hosts for sandboxed commands without prompting. Also raises nested-subagent depth to 3, adds MCP connection-failure diagnostics to claude mcp list//mcp, and ships assorted bug fixes (mid-stream API error text loss, permission approvals dropped on self-hosted runner restarts, Vim/screen-reader fixes).

Implications

  • Distrust-descends-the-stack / untrusted-config thread: strictAllowlist is another explicit-deny network control landing in Claude Code itself, continuing the pattern of sandboxing/trust primitives migrating into the harness rather than being left to ambient trust.
  • Capability clocks open-vs-closed thread: Opus 5 becoming the CC default (1M context, $10/$50 fast tier) is the closed-lab clock moving — this is the same release referenced in the 2026-07-25 daily journal entry as the “productized plateau” break.

← all signals