GPT-5.6 Sol vs Fable 5: Cost, Tokens, and Agentic Builds Head-to-Head

Key Takeaways

• Agentic builds — Fable wins quality, loses on cost. On a browser bike game and an interactive scroll-story site, Fable's output was clearly better, but a single Fable run cost ~$14–19 vs ~$1–4.50 for Sol. Fable emitted ~80–90k output tokens to Sol's ~20–31k.
• Token efficiency favors Sol/Codex. Sol consistently used far fewer tokens for a comparable task — though in these creative builds the cheaper output was also visibly weaker. The open question: if Sol were allowed to spend as much as Fable, would it match it?
• Freeform "five different things" test → Sol edged it on variety and polish (and looked "Claude-designed"), while being faster (7 min vs 15) and much cheaper.
• Stateless API one-offs → Sol "won" 24–3 across ~27 quick tasks — but largely because Fable often refused to answer (hard-baked safety guardrails). On score-when-answered they were near-even (Sol 0.98 vs Fable 0.966).
• Pricing: Sol runs roughly half Fable's input/output cost and lands close to Opus 4.8 — Nate argues Sol 5.6 is a fairer match for Opus 4.8 than for Fable 5.
• Mental model: Fable is the "manager / co-founder" — more creative, strategic, better at writing, brainstorming, and video, and it pushes back like a consultant. Sol is an excellent "worker" — cheaper, faster, better at computer use, and more careful at finding bugs / playing devil's advocate.
• Best of both: have Fable orchestrate a fleet of cheaper Sol worker agents.
• Caveat: Sol runs many more tests and sometimes over-thinks (especially Sol Ultra); Fable is less prone to over-engineering.

Commands & Code Mentioned

# Same goal prompt handed to both harnesses (Claude Code + Codex):
/goal build a genuinely fun, playable open-world bike game that runs in the browser