Single code-review at high effort using Fable 5.1 blows through entire session budget at $100/month subscription level without completing

**Bug Description**
The licensing model for Anthropic at the $100/month level is just horrible compared to OpenAI at the same level. With OpenAI, I no longer have session limits at that level - just weekly limits. I also can use as much Astra as I like. With Fable 5.1, attempting to do a single code-review at high level rapidly blows through my entire session budget without actually finishing the review and blows through my tokens with no benefit. This is a horrible customer experience.

This is the 3rd time this has happened to me recently, where my session usage is relatively low on the order of 20%. I attempt to get Fable 5.1 to do a `/code-review high` for something where about 1k lines have changed across about 15 to 20 files. Claude code rapidly spins up on the order of 8 sub-agents and blows through my entire current session budget without actually finishing the code review and shoots up my weekly usage for Fable by about 20% in the process without having achieved any useful work.

My experience with OpenAI and Astra in this regard is infinitely superior. Please fix this.

**Environment Info**
- Platform: darwin
- Terminal: tmux
- Version: 2.1.269
- Feedback ID: 9693665a-2d7b-47ec-a0e5-7e1a7d414002

**Errors**
```json
[]
```

For comparison, a single `/review` in Codex CLI using Astra at high on the identical code changes uses less than 1% of my weekly token budget at the $100/month subscription level. Something feels badly misconfigured or miscalibrated on your end.

NOTE: This same problem does **NOT** exist when using Opus 5 for `/code-review high` instead of using Fable 5.1, hence I do not believe the problem lies solely with the `code-review` skill.

Start by reproducing `/code-review high` on Fable 5.1 with about 1,000 changed lines across 15–20 files on darwin in tmux, then compare it with Opus 5. Inspect session and weekly usage during the run; done means the review completes without exhausting the session budget or consuming disproportionate tokens.

Rédigé par le modèle d'indexation à partir du texte de l'issue.