[BUG] Fable weekly usage counted (20%) with zero Fable requests since reset — only Opus 5.5 / Sonnet 5 used (Max 20x, Windows)

Preflight Checklist

• [ ] I have searched existing issues and this hasn't been reported yet — related: #96035 reports the same symptom (macOS, Opus 5, advisor off, no evidence attached yet). I'm filing separately because this report adds per-request evidence from local transcripts, and a different model (Opus 5.5) and platform (Windows).
• [x] This is a single bug report
• [x] I am using the latest version of Claude Code (2.1.284, same as npm latest)

What's Wrong?

Plan: Max (20x). Since my weekly limit reset on Sep 28, 8:59 PM (Asia/Seoul), I have not used a Fable model on any of my machines. I checked this machine's transcripts and my other machines. Even so, /usage shows Fable weekly usage at 20%:

Current session              2% used   Resets 10:49am (Asia/Seoul)
Current week (all models)   12% used   Resets Oct 5, 8:59pm (Asia/Seoul)
Current week (Fable)        20% used   Resets Oct 5, 8:59pm (Asia/Seoul)

The /usage session breakdown on this machine lists no Fable model:

Usage by model:
     claude-opus-5-5:  1.1m input, 3.1m output, 529.1m cache read, 11.4m cache write
     claude-sonnet-5:  1.4m input, 2.0m output, 334.0m cache read, 9.5m cache write
   claude-haiku-4-5:   124.2k input, 24.1k output, 1.5m cache read, 242.6k cache write, 6 web search

The Fable window is 50% of the weekly limit, so 20% of it is about 10% of the whole weekly limit. Roughly 10 of my 12 all-models points are being attributed to Fable, even though every request since the reset ran on a non-Fable model.

Per-request evidence.

• Method: I parsed every transcript JSONL under ~/.claude/projects modified since the reset: the main session, its subagents, and sessions spawned by plugins. I kept entries with timestamp >= 2026-09-28T12:00Z, deduplicated them by message.id + requestId, and grouped them by message.model.
• Result:

| model (message.model) | requests since reset | cache read | cache write | output | used by |
|---|---|---|---|---|---|
| claude-opus-5-5 | 295 | 64.9M | 1.54M | 312k | main session + subagents |
| claude-sonnet-5 | 149 | 14.3M | 0.73M | 122k | subagents |
| claude-opus-4-7 | 48 | 1.8M | 0.34M | 36k | automatic reviews by the security-guidance plugin |
| claude-fable-* | 0 | 0 | 0 | 0 | — |

(Uncached input is negligible for every model, under 1k tokens.)

• Other checks:
• In the same transcripts, every usage.iterations array contains exactly one {"type":"message"} iteration. No advisor iterations appear.
• server_tool_use.web_search_requests and server_tool_use.web_fetch_requests are 0 everywhere.
• The session has no advisor tool.
• Since the reset, the main session made 9 Agent (subagent) tool calls, and each one set model explicitly: sonnet ×6, opus ×3. None of them used fable. Resumed subagents keep their original model.
• Every request has usage.speed of standard or absent (fast mode was not used), service_tier: "standard", and inference_geo: "not_available".
• No settings file, environment variable (ANTHROPIC_MODEL, CLAUDE_CODE_SUBAGENT_MODEL, …) or agent definition selects Fable.
• Searching all transcripts in this project, at any date, finds zero claude-fable-* responses.

Workload: one long Claude Code session. The main model is Opus 5.5, and each request re-reads roughly 340k context tokens, almost all of them cache reads. The session dispatches subagents on Sonnet 5 and Opus 5.5.

Observation: Opus 5.5 is the largest share of my usage since the reset. The amount attributed to Fable (≈10% of the weekly limit) is close to what I would expect my Opus 5.5 usage alone to account for. So Opus 5.5 usage may be counted as Fable usage. I can't verify this from the client side.

What Should Happen?

The Fable weekly meter should only increase when a Fable model actually runs, whether as the main model, a subagent or the advisor. With zero Fable requests since the reset it should read 0%. Usage of Opus 5.5 / Sonnet 5 / Haiku should count only toward the all-models limit.

The help article (Claude Fable models on your plan) says: "You can use up to 50% of your weekly limit on Fable models, but your use of other models draws from the same usage limits". That doesn't suggest Opus/Sonnet usage raises the Fable meter. If it does by design, /usage and the docs should say so. If Claude Code calls Fable in the background for some reason, those calls should show up in /usage and in the transcripts.

Error Messages/Logs

No errors. The relevant /usage output is above.

Steps to Reproduce

1. After the weekly reset, use Claude Code only with Opus 5.5 as the main model (/model). Dispatch subagents on Sonnet 5 / Opus 5.5. Never select Fable, and don't use the advisor.
2. Work in a long session: in my case about 444 Opus 5.5 / Sonnet 5 requests since the reset, most of them large cached-context requests.
3. Run /usage. The "Current week (Fable)" meter is above 0% (20% here, i.e. ≈10% of the weekly limit).
4. Check the transcripts (message.model): there are no Fable requests.

Claude Model

Opus

Is this a regression?

I don't know

Claude Code Version

2.1.284 (Claude Code)

Platform

Anthropic API (Claude Max 20x subscription login)

Operating System

Windows (Windows 11 Pro 10.0.26200)

Terminal/Shell

Other

Additional Information

• Related: #96035 (same symptom, open), #80638 ("Opus use billed as Fable usage", closed as not planned), #97372 (desktop banner labels the Fable meter as the "Opus limit").
• Sample request IDs after the reset, in case they help trace how the usage was attributed: req_011CfVoW4KpApaHDKmCFgErB (Opus 5.5, main session, 2026-09-28T12:01Z), req_011CfVpTWr3EJcb3aE2L5ayZ (Sonnet 5, subagent, 12:14Z), req_011CfVuHEDxnvaSfqMacFP9C (Opus 4.7, plugin review, 13:17Z). I can provide more.
• I can share the small counting script (it only reads message.model, message.id, requestId, timestamp and usage from the local JSONL transcripts) if that helps.