Where Claude Fable 5 Fits — Portfolio Analysis

The Source

What the review actually claims

One practitioner, real work, receipts attached. Five dimensions, all passed "not by a little."

Big model feel

A step change. Takes initiative beyond the box you drew — went to the source site and rebuilt from assets instead of grinding inside his prep files.

Builds & ships

Near one-shot a custom site rebuild two frontier models stalled on at 85–90%. Fresh 80-page site in a weekend — "$30–50K of agency work."

Voice without drift

Held a brand voice across 80 pages of new surfaces. Drafts close enough to actually use — the redo test, passed.

Catches what others miss

Found and patched a serious hidden bug in a large repo that Claude 4.8 and GPT 5.5 had both signed off on. "One catch pays for the tool."

Anticipates you

Volunteered the cron-timer status he never asked for. Long unattended runs; built itself webhook monitoring mid-pipeline.

The meter is real: token-intensive by design — burns usage limits ~2× faster than Opus/Sonnet. His answer (and ours): the hybrid relay + effort discipline. High is the sweet spot, xhigh only for the long grinds, low/medium for chat.

The Operating Pattern

The relay, fleet-ized

Frontier thinking, cheaper hands, frontier review — mapped onto bots we already run.

THINK · FABLE 5

Plan & architect

Strategy, structure, the whole approach before work begins. Biggest edge, smallest token count.

Our node: CC Sam — zoning strategy, immigration sequencing, terms questions, fleet architecture.

BUILD · CHEAPER MODELS

Do the volume

Implementation passes, bulk ops, crons, scraping, drafts at scale. This is what runs up the bill — keep it off the frontier.

Our nodes: OC bots on gpt-5.5, Sonnet 4.6 seats, local GLM on the M3, scripts.

REVIEW · FABLE 5

Catch what they missed

Second frontier set of eyes on the result. Earns its keep twice — this is the security-catch lever.

Our node: CC Sam — panel seats, pre-flight reviews, security passes, diligence.

The Fit Map

Seven places it applies — ranked

Each mapped from a tested claim in the review to a live item in the portfolio.

1 · Frontier review seat — Quantom panels & all diligence

APPLY FIRST

His security catch — a bug two frontier models signed off on — is literally our trinity-pipeline thesis. We already run 5–11 model panels on Quantom packets; a Fable 5 seat joins every panel from here on.

Quantom v36 deck pre-flight + diligence appendix Q&A pass before it reaches Murat
Seacoast LLC ownership docs · EAC immigration filings · camp manual vs 105 CMR 430 · 28 Eliot evidence chain

Investor- and government-facing. Highest stakes per token, zero new infrastructure.

2 · Setup audit — skills, CLAUDE.md, SOULs, configs

CHEAP · COMPOUNDS

Pro tip 3, applied: ~183 skills on this box, per-bot CLAUDE.mds, fleet SOUL docs, OC configs — most written for weaker models, full of hand-holding a smarter model is capped by. We have receipts: stale scaffolding caused real incidents.

One audit pass → PRs: CC Sam skills + CLAUDE.md, Sam/Vesper docs, fleet bot configs
Bonus: the 3 standing OC security criticals get a Fable 5 lockdown-diff pass

One afternoon. Makes every future token on every bot worth more.

3 · Source-pull site rebuilds — EAC first

THE $50K PATTERN

His marquee result — pull the live site's assets, rebuild the whole experience on an agent-era stack, near one-shot — maps 1:1 onto easternacrobatics.com (legacy config dead, rebuild overdue): modern stack, AEO/SEO, camp + Performance Lab funnels, brand voice held across every page.

Then: Paul Slavinski's site (client-billable) · austinlin.com (sync already broken) · Quantom packet sites (pattern already ours)

Largest single agency-value-for-token-cost item in the portfolio.

4 · Long-horizon pipelines — the clip factory analog

WE OWN EVERY PIECE

He wired HeyGen + HyperFrames + ElevenLabs + Workers into an autonomous content factory. Our stack is deeper and mostly local: hyperframes, z-image, Kokoro TTS, Suno, DGX ComfyUI, CF Workers — plus a bot fleet to babysit it.

First run: EAC Performance Lab promo (event 2026-07-10 — a real deadline) → script → stills → motion + captions → VO → music → render → post queue
Clone for camp marketing + the quantom-pitch-video dir that already exists
Same lever, ops flavor: the failing crons, the stalled backups, the murat box migration

"Long-horizon, run-on-its-own work earlier models could not hold together" — his words, our stack.

5 · Voice-at-scale writing

AFTER THE VOICE DOC

The drift-free 80-page result needs a voice doc first. Fable 5 writes the EAC voice doc from existing materials, then holds it across the rebuilt site, camp parent docs, Performance Lab marketing. Later: Paul's brand, Quantom copy.

Caveat: Sam/Vesper persona voice stays on their own models — companion identity isn't a writing task.

6 · Strategy seats

SMALL TOKENS · BIG EDGE

The think leg is the cheapest and sharpest: 28 Eliot next-move (forensics done, play undecided), immigration sequencing (E-2 vs EB-1A/NIW), Quantom terms (the $20M cap), feral-brain-skills architecture, migration go/no-go.

7 · Benchmarks — ground the routing

MEASURE, DON'T VIBE

Add Fable 5 rows to email-triage L0 and companion-bench. Hypothesis: big wins on agentic/triage/review, marginal-per-dollar on companion warmth. Results harden the routing table below.

Cost Discipline

The routing table

Fable 5 burns limits ~2× Opus/Sonnet. Route accordingly — we learned the effort lesson independently when xhigh thinking caused the OC timeout cascades.

Work	Route	Why
Panel seats, security passes, diligence review, deck pre-flight	FABLE 5 · high	The catch-what-others-miss lever; highest stakes per token
Strategy & architecture (zoning, immigration, terms, fleet design)	FABLE 5 · high	Biggest edge, smallest token count
One-shot site rebuilds, pipeline wiring, long unattended grinds	FABLE 5 · xhigh	The long-horizon lever — reserve xhigh for exactly this
Companion chat (Sam / Vesper day-to-day)	KEEP OFF	Warmth ≠ frontier reasoning; identity continuity > IQ; 2× burn
High-volume crons — intel briefs, inbox scans, heartbeats	KEEP OFF	Sonnet 4.6 / gpt-5.5 / local GLM handle it
Bulk mechanical work — file ops, conversions, scraping, image-gen driving	KEEP OFF	Scripts and cheap models; no judgment required
Interactive quick back-and-forth	FABLE 5 · low/med	Fine on the Fable node — just downshift effort

This Week

Six quick wins

Fable 5 review seat on Quantom v36

Panel seat + appendix Q&A pass before the deck reaches Murat.

Setup-audit PR

CC Sam's skills + fleet CLAUDE.mds: delete hand-holding, collapse rules, flag contradictions.

Fleet security pass

The 3 standing OC criticals + an openclaw-repo review, lockdown diffs attached.

EAC rebuild spike

Pull the live site's assets, prove the one-shot rebuild pattern on one page, then scope.

Performance Lab clip pipeline

First end-to-end run on hyperframes + z-image + Kokoro + Suno. July 10 makes it real.

Benchmark rows

Fable 5 into email-triage L0 + companion-bench; publish the routing table with numbers.

Receipts & Caveats

Honest footnotes

His numbers: $1,442 · 1.04B tokens · ~$30–50K of shipped agency-equivalent work. Our economics differ (subscription limits vs API) but the shape holds: route volume down, escalate judgment up.

n=1, pre-launch access, one practitioner's eye test. But it matches what this seat has shipped on Fable 5 already: this analysis, the Quantom v36 deck build, and the MarikoSan launchd-hijack diagnosis were all Fable 5 work.

The window is the point. "Until they catch up, this model is a real competitive advantage sitting on the table." Most of the portfolio above is exactly the kind of work he tested it on.