I think I figured it out

Opus 5.5 for most work, Fable 5.2 for what can't be checked

Opus 5.5 is the model for daily work, and Fable 5.2 is the one to call when no objective test can judge the result. The subscription advantage holds, but the split is less tidy than the simple rule suggests. Fable's half-cost-per-task edge shrinks under subscription caps, and the verdict on Fable rests on about two days of use. [05:36]

Choosing between Opus 5.5 and Fable 5.2

1. Start most work on Opus 5.5 at medium effort. [07:35]
2. Raise Opus to high when medium stops checking what it should, and skip max, which overthinks simple tasks. [08:24]
3. Test the next model up at a lower effort level before paying for more reasoning past high. [12:20]
4. Switch to Fable 5.2 for hard, unverifiable work such as design, visual polish, or unconventional solutions. [36:11]
5. Move a stalled thread to Fable 5.2 when Opus loops, and let Fable judge whether the code is good enough to merge. [41:41]
6. Check your own weekly limit before leaning on Fable, since it can use only half of it. [16:42]
7. Use Haiku 5.5 for bulk classification and file scanning, not for writing code. [11:42]

Key topics

• Claude Opus 5.5 — The default model, strongest when a task can be verified and the spec is clear. Medium effort suits everyday work, but it ends turns early, so high is the safer setting for checking. Opus double-checks everything, which catches what Fable glosses over, and it orchestrates large migrations with sub-agents without trouble. It measures about 112 tokens per second. [18:19]

• Fable 5.2 — Suited to creative, exploratory, or unprovable problems, and it proposes ideas Opus never raises. It will also drop nitpicks and approve code, which makes it useful for merging stalled work. Its speed, around 106 tokens per second, matches Opus. In one evaluation of review quality it trailed Opus and Sonnet because it let slightly flawed code through. [39:56]

• Claude subscription limits — The $200 plan changes the economics. Fable is capped at half the weekly allowance, giving roughly three thousand dollars of Fable against fourteen thousand of Opus each month. The speaker reports that subscriptions beat every API alternative on tokens per dollar, a finding tied to one snapshot that could change. [03:22]

• Claude Sonnet 5.5 — Useful only as a tool for narrow, well-specified tasks. Without a clear path it runs long and costs about as much as Opus, so it has little place in the model picker. [10:13]

• Claude Haiku 5.5 — The fastest model, at about 255 tokens per second. It fits bulk pattern-matching across a codebase, though the speaker would never trust it with code. [11:24]