Source · release

Claude Code release 2.1.316

Summary

Claude Code 2.1.316 was released on 2026-10-31 (there was no public 2.1.315). It adds a setting for the effort level of subagents, a switch to turn off the ultrathink keyword, MCP form requests in cloud sessions, and several Claude apps gateway policy options. It also deprecates fast mode on Opus 4.8. The rest is fixes, several of them to the Bash tool's shell and to subagents.

Key points

  • Subagent effort: set CLAUDE_CODE_SUBAGENT_EFFORT (for example medium) to choose the effort subagents run at when nothing else names one. Before this they inherited the session's effort, so an xhigh session made every subagent expensive.
  • ultrathink off switch: CLAUDE_CODE_DISABLE_ULTRATHINK=1 stops the keyword from raising effort. In VS Code, ultrathink now runs its turn at xhigh, as in the terminal.
  • MCP forms in cloud sessions: when an MCP server asks for input (elicitation), cloud sessions now show a form card you can fill in or decline, instead of cancelling the request silently.
  • Fast mode on Opus 4.8 is deprecated as of 2026-10-30 and will be removed on 2026-11-30; after that, fast requests on Opus 4.8 fail rather than falling back. Use Opus 5.5 to keep the speed-up.
  • Gateway: pooled monthly budgets for groups (pooled_budget on the spend limits admin API), managed.policy_mode: layered so every matching policy applies, Bedrock guardrail input_tagging, and a check that every model a request names is allowed.
  • Fixes worth knowing: shell commands are no longer left running when Claude Code exits by itself; multi-line bash aliases and scripts with exported SHELLOPTS work in the Bash tool; subagents keep an inherited model when woken by a background job; --restricted sessions no longer upload files outside the working directories.
  • /code-review started from Desktop or the web now shows a findings card; /diff is cheaper on large repositories.
  • The Python Agent SDK 0.2.176 (2026-10-31) bundles this CLI version.

Why it matters

CLAUDE_CODE_SUBAGENT_EFFORT is a simple cost lever for multi-agent work: keep the main session at high effort and let subagents that search or summarise run lower.