Claude Code 2.1.267 adds a hard effort cap across providers

Claude Code 2.1.267 adds maxEffortLevel, a setting that caps reasoning effort on every request, including calls routed through Bedrock, Vertex and Foundry, according to the changelog. Set it top-level or per model inside modelSettings, and the cap holds as a ceiling. Requests can still run at a lower effort than the max, just never higher.

The more consequential fix in the same build is quieter: effort: frontmatter in custom commands, skills and subagents was being ignored on models with a pinned default effort level, specifically Opus 4.7, Opus 4.8 and Fable 5. If you wrote a skill file with effort: low and pointed it at one of those models before this release, the setting did nothing. The model ran at its pinned default the whole time.

That's a real cost problem for anyone who tuned effort per skill to control spend. A skill meant to run cheap and fast was silently running at full default effort instead, with no error and no warning in the output. The only way to have caught it was noticing the token bill didn't match the setting.

The release also adds --system-prompt-snapshot off, which re-renders the system prompt on every request instead of replaying the version recorded at the start of a conversation. That's useful while you're actively editing prompt text and want each turn to reflect the latest draft. It's the wrong setting to leave on afterward, because re-rendering breaks prefix cache stability, and caching is what keeps repeated system prompt tokens cheap.

Worth checking your own skill and subagent configs for effort: frontmatter if you're on a pinned-effort model. The bug had no visible symptom, so the only way to know if it hit you is to look.