I Made the Top Model My Session Default, and One Heavy Implementation Plus Its Review Burned Through Fable's Usage Limit

On August 22, 2026, the writer began a heavy implementation in a Fable session followed by a review. Unfortunately, Fable's usage limit was depleted almost immediately because the built-in skill responsible for the review inherited the session's model and executed on it. A few days later, the writer noticed a recurring issue in Opus sessions, which had picked up design work after the Fable usage limit ran out. These two incidents formed the primary focus of the story.

When the top model is set as the session default, the first thing to break is not the work product but the usage limit. Once the limit is depleted, judgment itself drops to a lower model. This article documents the paths the default leaked into, why existing conventions and warnings failed to prevent the issue, what ultimately stopped it, and the consequences once the limit was gone.

The writer uses Claude Code on a Max subscription plan, dividing the models across three tiers. Fable handles judgment, Opus focuses on implementation, and mechanical cross-checking is entrusted to Sonnet. The writer's settings.json file designates Fable as the default model, and he implemented a division of labor for handing implementation off in a previous article.

Since the beginning of August, every self-written subagent includes a model line. The setup included fable, opus, sonnet, and haiku agents, totaling 11 files in the ~/.claude/agents directory as of September 16, the day the writer began this account. All subagents are required to specify the model in their definition file, which the writer calls "pinning."

The default setting led to unforeseen consequences. Specifically, /code-review and /simplify, which come with Claude Code and require no model argument, ran on the session's model. When launched from a Fable session, they automatically used Fable, causing the usage limit to vanish in an instant. In addition to these two, there were other escape routes: the built-in subagents, which lacked a frontmatter.

The writer's log revealed that over three weeks, 576 agent launches included 373 built-ins, accounting for 65% of the launches. The log did not record the model, making it difficult to determine how many of these ran on Fable. This oversight was only discovered through meticulous counting and subsequent closure of most of the identified gaps.