The numbers
| Fable 5 | Opus 4.8 | |
|---|---|---|
| Model ID | claude-fable-5 | claude-opus-4-8 |
| Price (in/out, per MTok) | $10 / $50 | $5 / $25 |
| Context window | 1M tokens | 1M tokens |
| Max output | 128K | 128K |
| Thinking | Adaptive only; disabled 400s — omit instead | Adaptive only; explicit disabled accepted |
| Min cacheable prefix | 2,048 tokens | 4,096 tokens |
Where Fable earns the 2×
- Long-horizon autonomy. The headline gap. Stripe's 50M-line one-day migration and the FrontierCode top score are Fable results; with file memory, Fable improved 3× more than Opus 4.8 did on the same setup. Receipts on the evidence wall.
- Token efficiency. Fable is more token-efficient than past Claude models — Notation Capital measured a third of the reasoning tokens on physics research — so effective cost per completed task is closer than the sticker suggests.
- Vision. State-of-the-art; the Pokémon-FireRed-on-vision-alone result has no Opus equivalent.
Where Opus 4.8 stays the right call
- Daily-driver work. Opus 4.8 remains excellent for ordinary coding, writing, and analysis — at half the price. If a task wasn't straining Opus, Fable mostly buys you margin you don't need.
- High-volume production routes where cost dominates and tasks are well-specified.
- It's also Fable's own fallback — when Fable's safeguard classifiers trip (<5% of sessions), Opus 4.8 handles the response. Anthropic itself routes to Opus when Fable steps aside.
Migration cost: one line
Same API surface. Swap the model string; the only new 400 is explicit thinking: {"type": "disabled"} — omit the field instead. Full checklist in Recipe 01.
The routing rule
| Work | Send to |
|---|---|
| Overnight agentic runs, large migrations, hardest debugging | Fable 5 |
| Tasks where one wrong answer costs more than the tokens | Fable 5 |
| Daily coding, writing, review, analysis | Opus 4.8 |
| High-volume pipelines with tight budgets | Opus 4.8 (or Sonnet) |
Moral: don't ask which model is better — ask which tasks were failing. Those are Fable's.