# I Finally Get Fable 5 > [!question] Lingering Questions > - Does "high beats X-high/Max on cost-adjusted quality" hold outside Theo's stack, or is it specific to his repo size and task types? > - Where's the actual dividing line between "verification worth its token cost" and "verification more expensive than the work" — Theo hits the second case once but doesn't generalize a rule for predicting it in advance. > [!tip] Article Ideas > - Draft a `CLAUDE.md`-style glossary (reasoning effort vs. run length, staging-only merge gates) as a reusable agent-config template — Theo's video is the source, but the template should stand alone. > [!link] Cross Reference > - [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] > - Companion piece in this same batch, "A Field Guide to Fable: Finding Your Unknowns," makes the map/territory version of the same claim — this video is the concrete case where checking the territory (verification) got expensive. > [!quote] Vocabulary > - **Reasoning effort** (low/medium/high/X-high/Max/Ultra) — a per-step setting, not a run-length setting; Theo argues X-high and above over-reason and bloat diffs without buying proportional quality. ## Metadata - Source: [[3. Resources/Readwise/Articles/I Finally Get Fable 5]] - Author: Theo (t3.gg) - URL: https://www.youtube.com/watch?v=8GRmLR__OGQ ## Author's thesis Fable 5's value isn't per-step intelligence, it's autonomy horizon — how far it can run end-to-end through implementation, testing, verification, and multi-agent decomposition without stalling. Reusing Opus-era prompting habits underuses the model. Theo wants viewers to adopt cost-disciplined defaults (reasoning on high, not X-high/Max), let Fable orchestrate cheaper models for token-heavy subtasks, gate production behind a human, and verify hard — he shipped roughly a month of backlog in ~3 days for ~$150. ## Key arguments (≤3) 1. **Newer agent models run further, not just smarter** — the categorical shift is autonomy across implement/test/verify/sub-agent decomposition, not raw per-step reasoning quality. 2. **Cost discipline is a first-class design constraint**: reasoning effort is per-step, not per-run; higher settings (X-high/Max/Ultra) over-reason and bloat diffs without proportional quality gains. 3. **Verification is a real, sometimes dominant cost** — Theo burned more tokens verifying Fable's work in one case than Fable spent producing it, and the safety model that makes this tolerable is environment-scoped permissions (agents merge to staging, never to prod). ## Claim atoms ### Newer agent models run further, not just smarter > [!quote] ID:rw_fable5_h1 > Fable isn't just an improvement. Where/why: ~02:46, thesis of the video — the value shift is autonomy horizon (end-to-end implement/test/verify), not per-step intelligence; reusing Opus-era prompts underuses the model. ### Keep agent reasoning effort on high > [!quote] ID:rw_fable5_h2 > High is way smarter. Where/why: ~09:45, cost/quality tip — X-high/Max/Ultra over-reason and bloat diffs and overspend; effort is a per-step setting, not a run-length setting. ### Verifying agent work can cost more than doing it > [!quote] ID:rw_fable5_h3 > I burned way more tokens trying to verify the work that Fable did here than I burned getting Fable to do the work here. Where/why: ~42:46, post-merge stress test — verification is the safety budget in agentic workflows and should be treated as a first-class cost, not overhead. ### Agents may merge to staging, never to prod > [!quote] ID:rw_fable5_h4 > the model cannot ship or touch prod. Where/why: ~41:19 — explains why "merge with confidence" is tolerable: environment-scoped permissions (staging merges, human prod gate), not model trustworthiness alone. ## Permanent candidates - **A smart orchestrator should delegate to cheaper executors** (promoted — see linked Permanent note): Fable steering GPT-5.5 via Codex CLR for token-heavy/computer-use subtasks, evidenced by the ~$150 / 5.5-hour run. - The staging-only-merge / human-prod-gate pattern (H4) is a strong secondary candidate but overlaps with existing vault thinking on bounded action spaces for agent safety — link rather than duplicate. ## Processing checklist - [ ] Test the "high beats X-high/Max on cost-adjusted quality" claim on a real repo before adopting as personal policy - [ ] Add a staging-only-merge / human-prod-gate guardrail note to agent permission docs - [x] Connect into [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] ## Links - [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] — because applies-in: reasoning-effort tuning and verification cost are direct tokenomics decisions. - [[3. Resources/MOCs/AI Agent Safety MOC]] — because applies-in: the staging-only-merge / human-prod-gate pattern is a concrete bounded-action-space example.