A creator runs the same workout-app build through Opus 4.8, Fable 5, and three advisor-mode hybrids to find the cheapest way to get Fable-level output without paying full price.
Fable 5 produces the best-designed AI-built app, but running it as the sole builder is slow and can cost roughly $50 per project; using it only as a design advisor over a cheaper, subscription-included executor model gets close to that quality for near zero extra cost.
The host built the same 90-day workout-tracker app five ways to find the cheapest route to Fable 5-level design quality. Fable 5 alone produced the best UI but took 43 minutes and could cost up to $50 in API credits; Opus 4.8 alone was faster and cheaper but visually generic. Using Fable 5 as an on-call 'advisor' over an Opus 4.8 executor (Head Chef mode) was fastest of all and landed roughly halfway between the two in quality. A prompt-only method that just told Opus to 'think like Fable' produced broken, non-functional output. The winning combination swapped the executor from Opus to OpenAI's Codex under the same Fable-advisor pattern, producing the closest match to Fable's design sensibility while running on an existing subscription instead of paid credits.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →
Fable 5 is the best but most expensive model; the host sets up a test of cheaper ways to approximate its output.
A detailed single-file build spec for a 90-day 5x5 workout tracker ('NINETY') is sent to three parallel Claude Code sessions: Opus 4.8, Fable 5, and Opus 4.8 with Fable 5 as advisor.
Opus 4.8: 23 min. Fable 5 alone: 43:46 (up to $50 in API credits). Opus + Fable advisor: 16:29, the fastest of the three.
Clean, functional 'classic Opus' UI; workout flow, celebration animation, wall mode. Scored 6/10.
Visibly higher-class UI -- tighter spacing, nicer stopwatch and settings, top-aligned nav. The quality bar the rest of the video is chasing.
Opus 4.8 executing under a live Fable 5 advisor call. Better than Opus alone, still short of full Fable -- lands 'halfway between' the two.
A prompt/skill file telling Opus to imitate Fable's design process, with no live advisor call. Produces broken, non-functional UI -- called out as a failed approach.
A Peter Steinberger reply to the same advisor-pattern tweet suggests swapping the executor to OpenAI Codex. The host wires Claude Code to Codex under a Fable advisor.
The Codex-executed build is judged closest to Fable's design sensibility of any cost-saving option -- heavier typography, smoother chart, cleanest layout -- while running on an existing subscription. Video ends with a community-offer CTA.
Using an expensive, high-quality AI model only as a design advisor over a faster, cheaper executor model captures much of its polish without most of its cost.
“It's more of the head chef of the kitchen that delegates, looks at the menu, tweaks some things, tastes things with his pinky, but he's not doing the chopping, the stirring, and the cleaning up.”
“Fable actually took the longest at forty three minutes and forty six seconds, which if you're using API credits for this, it could be upwards of $50 for this one website.”
“It absolutely had an aneurysm and did not work.”
“This nice bar at the top that goes gradient colors, this to me by far wins and feels the most like Fable.”
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
Fable 5 is, by the host's own account, the best AI model for building well-designed apps -- and one of the most expensive to run outright. So he builds the exact same workout-tracker app five different ways, from full-price Fable to a broken prompt-only imitation, to find the cheapest setup that gets closest to Fable's output without the Fable-sized bill.
A workflow where an expensive, high-quality model is called in only for design decisions and hard problems while a cheaper model does the routine build work.
“Click the link in the description below and come and join us.”
Soft-launched at the very end after the technical content wraps, framed as ongoing personal help setting up AI workflows rather than a hard sales pitch.
00:00
00:08
00:19
00:25
00:32
00:4300:47
00:54
01:04
01:08
01:13
01:23
01:30
01:38
01:48
01:49
01:57
02:04
02:11
02:2102:25
02:36
02:4302:49
02:57
03:05
03:09
03:19
03:26
03:33
03:4003:46
03:52
04:03
04:10
04:17
04:21
04:2904:42
04:46
04:54
05:01
05:08
05:15
05:23
05:30
05:3905:44
05:48
05:59
06:06
06:11
06:19
06:25
06:35
06:43
06:4706:57
07:04
07:12
07:19
07:26
07:33
07:42
07:48
07:55
08:0208:08
08:17
08:26
08:32
08:35
08:46
08:53
09:01
09:08
09:15
09:22
09:3109:38
10:42A tweet-reaction breakdown of Anthropic's viral five-archetype framework — and the Slack-embedded Claude agent quietly filling the sixth role nobody named yet.
July 8th
12:53A YouTube host reverse-engineers Boris Cherny's one-line CLAUDE.md snippet, then runs a live three-way Claude Code test to prove a self-learning lessons file beats a static one.
July 5th
21:39How to point Karpathy's open-source AutoResearch at your cold emails, ads, and website so AI runs experiments while you sleep.
June 9th
30:40A 30-minute field report on burning $5,400 of subsidized AI inference in ten days — and what actually came out of it.
June 12th
13:34A 13-minute head-to-head where two Claude models race to clone the same landing page — one burns $30 and 35 minutes, the other $2.70 and 5, and the gap in quality tells you exactly when the expensive model earns its keep.
June 10th
1:02:24A senior developer's real AI-agent setup, and the argument that the harness — not the model — is where the leverage lives.
June 18th