Model feedback: Fable 5 missed 2 errors in a ~10M-token project (caught by Grok 4.5 cross-review); pricing will decide model choice

Filed by Claude Code on the user's explicit request — user-relayed model feedback for the Fable 5 team.

What happened

• Ran a ~10M-token project overnight on claude-fable-5 (Claude Code CLI).
• Pasted the finished deliverable into Grok 4.5 for an independent review.
• Grok 4.5 found two real mistakes that Fable 5 had made and not caught.

User's position (verbatim intent)

"Ran a 10M token project last night on Fable, pasted it over to Grok 4.5 — it found two mistakes. When Fable is on token-only [billing], I will be on Grok 4.5."

i.e. as a heavy long-context user, the moment Fable is only available on per-token pricing, model choice becomes a pure price/quality comparison — and cross-model review results like this one currently tip that comparison away from Fable.

Asks

1. Long-context (multi-million-token) error accumulation: treat as a quality priority; adversarial self-verification on large deliverables would have caught these.
2. Pricing/packaging for heavy users matters as much as capability — please factor this into how Fable is offered.

Environment: Claude Code CLI on macOS (darwin 25.2.0), model claude-fable-5.