firehose> #llmops
video youtu.be Watch ↗

Fable 5 + GPT 5.6 Sol = CHEAT CODE

TL;DR

Chase AI's argument is that "which model is better" is the wrong question, and he ships a Claude Code skill family (/grill-me-codex, /codex-build) that runs a four-stage pipeline instead: Fable interviews the operator (Matt Pocock's grill-me), Fable and OpenAI Codex argue the plan to consensus over a capped number of rounds, Codex writes the code from the frozen plan, and Fable reviews the diff like a contributor PR — with a bounded two-round repair loop before Fable takes the wheel itself. The load-bearing idea is not the token saving the title advertises but the independence claim the skill's own README makes: the same model that plans the build and writes the build can't be trusted to grade its own work — it's an echo chamber; a different provider catches what Claude structurally can't see in itself. That is a direct, on-the-record answer to the open question this vault already carries on Adversarial Planning Council — whether persona diversity on one model substitutes for genuine model diversity — and it comes down hard on the model-diversity side, with the cross-vendor role swap (nobody grades their own work, in both directions) as the mechanism. The economics are the secondary argument and the weaker one: the source asserts GPT-5.6 Sol benchmarks ahead of Claude Fable 5 and is cheaper than Opus, so the executor slot in a planner/executor split should be filled by a rival vendor's model rather than by a smaller model on your own vendor's ladder — which is where the source explicitly breaks with Advisor Mode, the pattern he himself put in this vault.

Concepts introduced

Held, not dropped — themes this capture touches that do not warrant a concept page yet:

Key claims

Why this refines the graph

On Adversarial Planning Council — this source closes an open question the page has been carrying. That page records a live tension: Karpathy's original council polled different LLMs, but llm-council-skill runs five personas on a single model and its author reports it "still works quite well" — logged as "asserted from use, not measured — a testable refinement, not a settled result." The page then sharpens the doubt itself: an author-constructed panel on one model with a shared framing yields correlated, not independent judgments, so agreement across lenses is "a strong hypothesis, not independent proof." This source takes the other horn and states the reason plainly: the echo chamber isn't a prompt problem, it's a lineage problem — "a different provider catches what Claude structurally can't see in itself." It does not settle the question empirically (no measurement is offered either way), but it converts the vault's suspicion into a named counter-position from a practitioner who shipped on it, and it supplies the mechanism the persona-only council lacks: the critic is a different vendor's model, so its blind spots are uncorrelated by construction rather than by instruction. Two related refinements come with it: the council here runs on two members rather than a roster of five (see Bounded Fan-Out — model diversity may buy with two what persona diversity needs five to approximate), and the council's verdict vocabulary shifts again — not Go/Reshape/Kill, but converge-or-cap.

On Model-Tier Routing — the second tier here isn't a lower tier. That page already holds cross-vendor delegation as a routing target ("/codex:rescue, /codex:transfer to delegate to GPT-5.5"), but frames the split as capability tiers: premium model plans, cheaper-and-dumber model executes. This source routes across vendors to a model it claims is not weaker — the argument is that the executor slot should go to a peer on a different meter, chosen for independence and for which allowance it burns. That is a genuinely different selection criterion than "cheapest tier that clears the bar," and it stacks with the page's existing "expensive boss that never codes" org-chart framing rather than replacing it.

On Advisor Mode — a tension inside the vault, from the source that put the concept here. Advisor Mode was distilled from make-fable-5-80-cheaper-other-usage-cheat-codes — the same channel. That page holds: a cheap executor consulting a smarter advisor when stuck beats the cheap model solo on both score and cost (Sonnet 4.6 High + Opus advisor, 74.8% @ $0.96 vs 72.1% @ $1.09). This video asserts the opposite recommendation: "I think this is way better than passing things off to Opus or to Sonnet or using advisor mode inside of Claude Code, because these GPT models are just better than those smaller Anthropic models and they are cheaper." The tension to resolve is narrow and worth stating precisely: the video does not dispute advisor mode's measured result, and does not dispute the planner/executor shape — it disputes the executor selection, arguing the in-vendor ladder is dominated by a cross-vendor option. Both positions can hold simultaneously only for an operator paying for both vendors; for a Claude-only operator, Advisor Mode remains the live option and this video's route is unavailable. Neither side is adjudicated here; the deciding evidence would be the benchmark claims above, which are vendor-published and unverified. Note also that the two sources are not independent — same channel, ten days apart — so this is one practitioner revising his own recommendation, not two sources disagreeing.

On Spec-Driven Development — corroboration, not news. The grill→plan→implement→review pipeline and the "plan mode on steroids" interview are already on that page as a fixed pipeline of named skills, credited to the same upstream author (Matt Pocock). This source is a third instance of front-loading attention onto the plan; it adds the cross-model hardening step between spec and implementation, and nothing else new.

Illustrated walkthrough

t=00:57 — the benchmark case, hedged in the same breath. On screen: openai.com/index/previewing-gpt-5-6-sol/, a TerminalBench 2.1 bar chart reading GPT-5.6 Sol Ultra 91.9%, GPT-5.6 Sol 88.8%, Claude Mythos 5 88.0%, GPT-5.6 Terra 84.3%, Claude Fable 5 84.3%, GPT-5.5 83.4%, GPT-5.6 Luna 82.5%, Claude Opus 4.8 78.9%, Gemini 3.1 Pro Preview 70.7%. Said over it: "Soul 5.6 is wildly powerful, at least according to the benchmarks. Now, grain of salt, this is coming from OpenAI." The source flags its own vendor-published evidence — worth preserving, since the whole cost argument rests on this chart.

t=01:49 → 02:02 — the token-efficiency curve. A GeneBench v1 score-vs-API-cost plot; the hover tooltip at 02:02 reads Model: GPT-5.5 · Reasoning effort: Xhigh · API Cost (USD): $1.24 · Score: 22.94%. Narration rounds this to "23% at $1.24" and contrasts "5.6 is a 25% score… at 56 [cents]. So way cheaper and therefore way more efficient."

t=02:20 — the frame that says slightly more than the narration. On screen: deepswe.datacurve.ai, DeepSWE score against avg cost per task. Narration: "direct comparisons of 5.5 versus Opus 4.8 — there's really no contest. Higher pass rates, lower cost." The plotted curves do show gpt-5.5 [medium] above claude-opus-4.8 [high], supporting that specific claim — and the same chart also plots claude-fable-5 [high] above both, at ~68–70%, which the narration does not mention. The source's argument is about the executor slot, where Fable is not the candidate; noted because the visual channel carries a comparison the audio doesn't.

t=03:00 → 04:30 — the pipeline, drawn. An Excalidraw board titled /codex-build & /grill-me-codex, built up across three sampled frames into: 1) Interview [grill-me] ← annotated Fable; 2) Adversarial planning ← annotated F ⇄ Codex (arrows both ways); 3) Codex builds; 4) Fable Reviews (an arrow loops back from 4 to 3). Spoken alongside: Fable is "driving the ship"; Fable and Codex "go back and forth for a maximum of five iterations… till they reach a consensus"; and at review time Fable will tell Codex "you did that wrong, fix it" — "it's going to do that twice. If by the third time it's not complete, well then Fable will clean it up itself." The 60 s hole sits inside this stretch.

t=05:21 — the run starts, and the meter is visible. A Claude Code v2.1.204 session: Fable 5 with high effort · Claude Max, and a banner reading "Extended: Fable 5 is included in your weekly limit — Through July 12, you can use up to 50% of your weekly usage limit on Fable 5… Fable 5 draws down usage faster than Opus 4.8." Below it: /grill-me-codex Build "Trip Atlas" — a stylized, cinematic trip planner web app. That banner is the unstated premise of the entire video: the thing being conserved is a weekly allowance, not a dollar bill.

t=06:00 — Act 1 interviews, and recommends. Grill-me-codex Act 1 start… Will interrogate one question at a time, recommend an answer each time. Q1 — What is this FOR? with three options (video demo / real personal tool / shippable product) and, crucially, My recommendation: A with B's bones. Your pattern (localflow, chomp, jarvis-hud) = build real working thing, video is the payoff. The interview isn't neutral elicitation — it reads the operator's own history and argues for an answer, then asks "Which is it?"

t=07:00 — consensus, logged. Converged: APPROVED, 2 rounds. Artifacts at ~/projects/trip-atlas/ — PLAN.md (final) + PLAN-REVIEW-LOG.md (full argument). The summary: Act 1 locked identity; Act 2 round 1 (Codex, 12 findings) hardened the data core — stable geonameId round-tripping, civil-date strings killing TZ off-by-ones, antimeridian splitting, a replay-vs-edit race fixed via trip snapshot, pins moved into SVG ("killed whole DOM-sync problem class") — 11 accepted, 1 scoped down with logged reason; Act 2 round 2 approved with 3 non-blocking notes folded in. Brewed for 8m 14s. Then the fork: "Grilled + survived 2 rounds of Codex. Implement now — Codex builds it (/codex-build), Claude builds it, or stop here?" Two details do real work here: the argument is written to a file (PLAN-REVIEW-LOG.md), and disagreement is resolved and recorded (11 accepted, 1 rejected with a logged reason) rather than deferred.

t=07:30 — the handoff. Skill(codex-build) — Successfully loaded skill.

t=08:30 → 09:00 — the artifact. Trip Atlas at 127.0.0.1:5173: "The Brass Compass Route", a parchment vector world map with numbered pins, a stops panel (Lisbon / Istanbul / …), 3 stops · 17 days · 6,680 km, an Offline DB marker on the city lookup, and a Cinematic Replay mode. Verdict spoken: "not bad, I think, for the first pass… what this really was about was just showing this workflow in action." Then the number the title is selling: "we only burned up about 130,000 tokens on the Fable side to get this whole thing done."

t=09:30 — the README states the real thesis. github.com/chaseai-yt/grill-me-codex (378 stars, 41 forks): "Two AI models harden your plan — then swap jobs to build it.… Act 3 (optional) flips the roles: Codex writes the code from the frozen plan while Claude reviews the diff like a contributor PR. Cross-model checks in both directions — nobody grades their own work." And the block quote: "Why a second model? Because the same model that plans the build and writes the build can't be trusted to grade its own work — it's an echo chamber. A different provider catches what Claude structurally can't see in itself." Credited: built on Matt Pocock's grill-me / grill-with-docs (MIT, Act 1 is his work); Act 3's delegation pattern adapted from Peter Steinberger's codex-first.

Linked from