Session-wide model switch to Fable billed all turns at premium rate — no warning when scoping was intended to be per-task
Summary
I intended to use Fable (claude-fable-5) as a per-task/per-agent model override for a small
number of high-context orchestration dispatches — not as my interactive session's overall
default model. Instead, my interactive session itself ended up running on Fable end-to-end,
which billed every turn (routine status checks, low-complexity orchestration, etc.) at
Fable's higher per-token rate instead of only the intended dispatches. This burned through
my usage budget much faster than expected (down to ~79% remaining well ahead of my normal
reset window).
What I expected
Switching models for a specific task/dispatch (e.g. via a subagent call or a one-offclaude --model claude-fable-5 invocation) to stay scoped to that dispatch, while my
interactive session continued running at its normal default tier (Sonnet/Opus).
What happened
The interactive session's default model was set to Fable, so all subsequent turns in that
session — including ones that had nothing to do with the Fable-worthy task — were billed at
Fable's rate. There was no warning or confirmation step distinguishing "switch this whole
session to a premium/routing-only tier" from "route one task to it."
Ask
- Consider a confirmation/warning when
/model(or equivalent) is used to set a
routing-tier/premium model as the session default, similar in spirit to a cost-guard —
at minimum, docs/UI language that makes clear a session-wide model switch applies to
every subsequent turn, not just one dispatch.
- Separately, I'd like the affected usage reviewed for a possible credit — is there a
supported channel for that, or should this go through Anthropic Console billing support
instead of a GitHub issue?
Environment
- Claude Code CLI, models involved:
claude-fable-5(session default, unintended) and
claude-sonnet-5 (reverted to, intended default)
- Reported same-day as a workstation battery failure that prompted a usage review
This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗