Billing/routing: /model set to Opus 4.8 but billed to Fable 5; secondary model's 100% limit gated a 56h session

Claude Code — billing/routing mismatch: selected model (Opus 4.8) but billed to Fable 5; a secondary model's 100% limit gated an entire 56-hour session

Product: Claude Code CLI
Version: 2.1.201
Date of failure: 2026-07-05 (UTC)
No chat content included — model/billing facts only.

Summary

The model actually served and billed diverged from the model selected in /model. /model was set to and kept on Opus 4.8 (claude-opus-4-8), but the backend served the majority of the session as Fable 5 (claude-fable-5). When Fable 5 reached its 100% limit, that limit gated the entire session — every subsequent generation returned only the Fable 5 limit message, and switching models via /model did not restore service (each selection produced exactly one response: the same Fable 5 limit notice). This froze a long-running session and its in-progress work could not be saved.

Evidence (from the session transcript .jsonl, message.model field)

Actual model that produced each assistant response over the session:

| model id served/billed | assistant responses |
|:--|--:|
| claude-fable-5 | 694 |
| claude-opus-4-8 | 348 |
| sonnet (subagents) | 13 |
| opus / <synthetic> | 2 / 7 |

• The selected model in /model was Opus 4.8, kept throughout — yet the served model silently alternated between claude-opus-4-8 and claude-fable-5 across the session (not a single, stable selection).
• Session length: ~56.5 hours (2026-07-02T22:22:21Z → 2026-07-05T06:54:35Z; 2007 timestamped transcript entries).
• Failure sequence (2026-07-05, UTC):
• 06:49:40 — You've reached your Fable 5 limit. Run /usage-credits to continue or switch models with /model.
• Each subsequent generation returned the same Fable 5 limit message.
• /model was re-invoked (06:51, 06:54); each attempt still bounced off the Fable 5 100% limit — no selected model could run.
• Only reliable model verification was the message.model field in the transcript; the /model status line showed Opus 4.8 while the backend billed Fable 5.
• Workaround found: a full CLI restart cleared it (a fresh session runs cleanly on claude-opus-4-8, no Fable). In-session /model switching did not fix routing.

Impact

• Weekly usage (all models) is now at ~90%, session at 100% — consumed in large part by Fable 5, a model that was not intentionally selected and whose billing the user did not confirm.
• A ~56-hour session was rendered unusable at the point of the Fable 5 cap, with no way to persist development-stage progress via model switching.

Requests

1. Fix routing/billing so the served model matches the /model selection, and so a secondary model's usage limit cannot gate a session whose selected model still has quota.
2. Reset the weekly limit for this account, since the Fable 5 consumption was not user-selected or user-confirmed (the account was billed for a model other than the one chosen in /model).

Account

Account identifiers and the full session transcript (model IDs + timestamps; chat content withheld) can be provided privately to Anthropic support on request.