CONFIRMED. Claude Code 2.1.267 introduces a change that looks small but affects the core of any development agent: teams can now cap the maximum effort level applied by each model, including executions routed through Bedrock, Vertex, or Foundry. The release also lets users disable the system prompt snapshot, causing the system text to be rendered again on every request. In practice, Anthropic is adding more operational control and a cleaner way to iterate on prompts while closing a long list of failures.
The maxEffortLevel setting can be defined globally or per model under modelSettings. Users can still select a lower effort level, but cannot exceed the configured ceiling. This matters to teams running agents across hybrid environments because the limit follows the workflow even when the provider changes. For a small Brazilian company, the benefit is not automatically “more intelligence.” It is predictability: a simple task should not jump to the most expensive and slowest mode merely because the model or gateway changed.
The --system-prompt-snapshot off option targets a different but very practical problem. Claude Code normally reuses the prompt recorded in the conversation. Turning the behavior off lets people maintaining commands, skills, or subagents test prompt changes without opening a new session. That shortens iteration cycles, but it also increases the need to version and review prompts. A system that changes its system text during a task can move faster while becoming less reproducible.
The rest of the release is less visible, but perhaps more important in daily use. It fixes Cowork scheduled tasks that failed when organization policies required sandboxing, improves recovery for large sessions, and prevents parallel calls or hook output from disappearing after a resume. It also addresses reconnection problems in tmux and SSH, MCP tools added mid-session, prompt caching, expired AWS and Google Cloud credentials, and Remote Control clients.
Several security fixes deserve special attention. Marketplace entries containing a backslash can no longer bypass the containment check on macOS and Linux. Unreadable managed settings for hook URLs, environment variables, and channel plugins now admit nothing instead of everything. These fixes do not make for an impressive demo, but they reduce the chance that a corporate policy fails open.
My verdict: update development environments and test before broad production rollout. This release does not turn Claude Code into a different product, but it improves three properties that determine whether an agent is operable: controllable cost, iterable prompts, and reliable resumption. For small and midsize companies, the priority is to validate five real workflows, including a long MCP session, a resume after compaction, and a sandboxed scheduled task. The marketing is in the new knobs; the real value is less surprising behavior.
Sources
Official GitHub release: https://github.com/anthropics/claude-code/releases/tag/v2.1.267
Structured release record: https://api.github.com/repos/anthropics/claude-code/releases/tags/v2.1.267
Official effort configuration documentation: https://code.claude.com/docs/en/model-config
Independent update history: https://releasebot.io/updates/anthropic/claude-code