TIC·TAC·TOE
←↑↓→ move · Enter/Space place · N new round
Build Log — Opus 4.8 vs Fable 5
Same prompt, two Claude models. This game was built by Claude Opus 4.8; a sibling build, GRIDLOCK, by Claude Fable 5 — same effort level (ultracode). Each was audited by a 3-agent adversarial workflow. The honest head-to-head:
It’s close. Fable’s first-pass code was actually cleaner — it avoided a class of bug Opus had to catch in review. But Opus met the constraints that defined the task — truly offline / self-contained, with committed proof — that Fable missed. Opus wins the spec; Fable is the prettier artifact.
| Metric | Opus | Fable |
|---|---|---|
| AI unbeatable | ✓ proven | ✓ proven |
| External requests | 0 | 3 (Google Fonts) |
| Works offline | ✓ fully | fallback only |
| Tests committed | ✓ runs | ✗ claimed, absent |
| Retina canvas | ✓ sharp | ✗ blurry |
Where Fable won
- Cleaner first pass — no AI-timer races or focus bugs (Opus fixed those in review)
- Richer feedback: deny-flash + “cell already marked” on illegal moves
- Stronger visual identity: custom HUD fonts, notched console panels, randomized AI play
Where Opus won
- Zero network requests — genuinely self-contained & offline (the actual spec)
- Exhaustive AI proof committed & runnable (Fable claimed tests, shipped none)
- Screen-reader win state, backdrop-filter fallbacks, DPR-correct rendering
Unbeatability proven by exhaustive search — Opus 569 games, Fable 6,112 — both with 0 human wins. Source: tic-toc-opus · gridlock.