How Setting a Top AI Model as Session Default Silently Drained Usage Limits

On August 22, 2026, a developer using Claude Code on a Max subscription plan exhausted their Fable model's usage limit after running a heavy implementation and a code review in the same session. The issue arose because built-in tools like /code-review and /simplify inherit the session's default model, meaning they ran on Fable — the highest-tier model — rather than a lower one. Further investigation revealed that built-in subagents, which lack model-pinning frontmatter, accounted for roughly 65% of all agent launches over a three-week period, creating a widespread model-leak risk. Once Fable's limit was exhausted, judgment tasks defaulted to a lower model, leading to over-engineered outputs observed in subsequent Opus sessions. The developer documented the leak paths, explained why existing conventions and warnings failed to prevent the issue, and outlined the configuration changes — including setting the CLAUDE_CODE_SUBAGENT_MODEL environment variable — that ultimately closed most of the gaps.