Vol. 170 Fable 5 重出江湖，GPT 仍需努力
Summary
This 枫言枫语 episode by Justin Yan and 自立 uses the reopened Fable 5 release to compare frontier coding-model capability, Codex execution, Superpowers orchestration, and GrillMe Skills-style manual skill workflows. The hosts argue that Fable 5 feels unusually strong at planning, requirement clarification, one-shot implementation, review triage, and UI generation, but its value is constrained by quota, API cost, and the need for human taste and verification. The second half extends the coding workflow discussion into Token-Driven Software, dynamic generated interfaces, AI-native games, independent developers, and the commercialization difficulty created by AI Inference Cost Structure.
Key Claims
- Fable 5 is described as possibly weaker than its earliest briefly-opened version, but still stronger in the hosts’ practical coding workflow than Opus 4.8 and GPT 5.5.
- One-Shot AI Coding improves when the model can decompose complex requirements, generate usable UI, anticipate edge cases, and leave only small review issues rather than P0 defects.
- The preferred workflow is to let Fable 5 discuss requirements, write plans, PRDs, or issues, then let Codex execute or review implementation.
- Superpowers is useful for non-experts because it enforces brainstorming, specification, planning, subagents, TDD, review, and acceptance, but it can overcomplicate small tasks and burn a large amount of tokens.
- GrillMe Skills are presented as a lighter alternative for experienced users: manually invoke requirement questioning, ADR/spec/PRD generation, and issue decomposition only when the task warrants it.
- Long-running automation still needs AI Coding Verification: the model can drift, and the workflow needs step-by-step acceptance, tests, review, and ways to pull the agent back to the goal.
- The episode reinforces AI Inference Cost Structure and AI Subscription Economics through Fable-specific limits, faster quota burn, and a small API change reportedly costing about five dollars.
- Token-Driven Software imagines software whose interface, branch logic, NPCs, AR effects, or interaction flow can be generated from user context instead of fixed in advance.
- AI may lower the production cost of independent apps, games, and short-form content, but “usable” output is not the same as elegant, differentiated, or defensible product quality.
- 2C AI products face a harder payment problem than many 2B workflows because high token cost makes unlimited free usage difficult while large companies can subsidize usage longer.
- Model Routing Cost Control becomes necessary when simple tasks can use cheap models while complex planning, review, or product judgment should route to more capable models.
Key Quotes
“满血版” — the hosts’ term for the stronger initial Fable 5 release they believe may have been briefly available.
“能用” and “特别优雅” — the distinction used to separate generated working tools from products with taste and polish.
“token 已经像一种底层资源” — the closing cost-control frame for choosing models by task value and difficulty.
Connections
- 枫言枫语, Justin Yan, and 自立 — show and host context.
- Fable 5 — central model/product discussed through coding capability, quotas, and review quality.
- Codex, AI Coding Verification, AI Engineering Thinking, and Vibe Coding — practical coding workflow and verification layer.
- Superpowers, GrillMe Skills, AI Skills, Agent Harness, and Subagent Workflow — orchestration and skill-selection layer.
- AI Inference Cost Structure, AI Subscription Economics, and Model Routing Cost Control — token cost, limits, and model-routing layer.
- Token-Driven Software, On-Demand Apps, Generated Work Interfaces, and Agent Native Software — software-shape and interface speculation.
- AI Interactive Entertainment, World Models, and Product Led Willingness To Pay — game, world-simulation, and commercialization extensions.
- Anthropic, OpenAI, Google, Gemini, and DeepSeek — model-provider and competitive comparison points.