I Made Opus 4.8 and Fable 5 Build the Same App (RAW RESULTS)

Summary (Technology/Product Comparison)

The creator runs head-to-head tests between Opus 4.8 (their build is “Opus48”) and Claude’s flagship model, “Fable 5”.

They use:

• The same overall prompt
• A one-shot generation rule: no revisions, single output
• A live deployment approach: the resulting apps are run in the browser

The video is structured into multiple “builds,” each stress-testing different capabilities:

• UI completeness
• Agent orchestration
• Data extraction
• 3D realism
• Interactive behavior

A custom usage/cost tracker estimates input/output tokens and usage-based pricing (noting Fable 5 pricing would become usage-based soon). Build times and token efficiency are compared.

---

Build 1: Full E-commerce Store (“Slowburn”)

Goal

Build a Shopify-like e-commerce site with:

• Landing page + catalog + full architecture + database
• 30 distinct products
• Visually distinct product images
• Readable text (no low-contrast typography)
• Avoid common failure modes (e.g., missing labels, missing hero image, identical products)

Results

Opus 4.8 Output (Store)

• Some improvement over a prior attempt (“dynamic workflows”), including readable text and labels.
• However, there were noticeable UX/UI issues:
  • Product text/layout readability problems (e.g., candle name placement)
  • Weird CTA button
  • Category filters felt incoherent/overly scattered
  • Confusing cart icon appearance
  • Navbar UX described as “terrible UX”

Fable 5 Output (Store)

• Cleaner, more complete e-commerce UX:
  • Cleaner header/hero area
  • Better buttons/notifications
  • Much better filtering categories, e.g. “Shop by ritual” (morning ritual / deep work / unwind / sleep)
  • Overall layout/details feel more “Shopify-like”
• Stronger integration with image generation:
  • Effective prompting for image generation
  • Consistent aesthetic (e.g., candle look)
  • Varied imagery with better colors/lighting and subtle polish

Conclusion (Build 1)

Fable 5 is the clear winner for store UI/UX quality and completeness.

Cost & Speed Notes (Usage Estimate)

• Opus 4.8: ~198k output tokens, estimated $21.41, and slower overall (later framed closer to ~50 minutes).
• Fable 5: A place listed ~518k output tokens, but later comparisons indicate it’s more token efficient with lower estimated total spend and ~15 minutes faster for the same build.

Note: The transcript’s numbers are somewhat confusing, but the consistent takeaway is: Fable 5 was higher quality and faster, with usage-based cost remaining non-trivial but reasonable for the scope.

---

Build 2: Interactive 3D Art Museum of Art History

Goal

Create a browser-based interactive, zoomable infinite-canvas timeline of major art periods where users can:

• Zoom into art movements/periods
• Click artists for museum placards
• Enter a 3D gallery to view extracted paintings
• Pull data from Wikipedia (Wikipedia API/comments are mentioned)
• Store/organize into a database (referred to as “Neon”)
• Emphasize realism: “not a lazy 3JS demo” (lighting, tone mapping, shading, etc.)

Important Constraints / Issues

• A public-domain vs modern art mismatch is flagged.
• Picasso/Dali-style cases introduce copyright concerns.
• The creator adjusts scope to use public-domain images for the demo.

Results (Functionality & Agent Coordination)

Fable 5

• Produced a sophisticated timeline with color coding and smooth zoom.
• Demonstrated strong interactivity:
  • Click animations (GSAP referenced)
  • Navigation remains usable while interacting
  • Fluid loading/movement for artists and galleries for a “real museum” feel
• Reported breakdown/status:
  • 16 periods
  • 69 hours (likely process time, presented as a status figure)
  • 767 paintings in the Neon database
• Strong signs of agent orchestration:
  • Parallel work
  • Queued status updates
  • Reinvocation when tasks finish

Opus 4.8

• Timeline view worked reasonably (color-coded, zoomable canvas behavior).
• But clicking into galleries failed:
  • Interaction/canvas movement overlaps with click targeting
  • Result described as: “You can’t click in anymore”
• Hierarchy/organization was praised, but the key requirement—entering galleries—didn’t work reliably.

Conclusion (Build 2)

Fable 5 wins due to:

• Better interactivity (click-to-enter galleries works)
• More realistic/immersive 3D environment
• More reliable navigation and UX

---

Build 3: Age of Empires-Style Real-Time Strategy (3D)

Goal

A playable browser RTS with:

• Full 3D world (WebGL/3JS equivalent)
• Core loop from the town center
• Units, enemies, civ/build mechanics
• Painterly/realistic materials and lighting (explicit rejection of low-poly stylization)

Results

Opus 4.8

• Visual output far below expectations (“blobs”).
• Gameplay/app felt broken or non-functional:
  • Couldn’t properly navigate/move the camera
  • Unclear controls and inability to operate the world

Fable 5

• Much higher graphics fidelity—described as nearly matching Age of Empires-like visuals.
• UI/control loop felt coherent:
  • Map movement and selection interactions worked (left/right click described)
  • Building placements (farm, house)
  • Enemies visible with armor and combat-related structures
• Creator suggests it appears fully playable, even mentioning an intention like “available on Steam” with a price shown (299), though it may relate to the demo/plan mentioned in the transcript.

Conclusion (Build 3)

Fable 5 is the clear winner for the RTS task:

• Rendering quality
• Interaction reliability
• Perceived game completeness

---

Overall Analysis / Theme

Across all three builds, the creator repeatedly finds:

• Fable 5 produces higher-quality, more complete, and more interactive applications
• Opus 4.8 often fails on critical UX/interaction reliability (e.g., gallery clicks, RTS navigation)
• Fable 5 is generally faster and more effective in these one-shot deployments
• Image generation integration is a major advantage for Fable 5 in these tests

---

Main Speakers / Sources

• Main speaker: the YouTube creator/host (unnamed in subtitles), who runs the experiments in Cursor
• Model references:
  • Opus 4.8
  • Claude “Fable 5”
• Other referenced tools/inputs:
  • GPT-image / image generation behavior (referred to as “GB image 2” in subtitles)
  • Wikipedia API (art museum data source)
  • Neon (database mentioned)
  • 3JS/WebGL (interactive 3D experiences)
  • GSAP (animation library referenced)