Fable 5.2 won four creative tests; Opus resisted pressure better

SUMMARY

In 30 seconds

Peter Yang compared Fable 5.2 and Opus 5.5 on video, games, website design, writing and whether they would challenge a flawed business plan. He preferred Fable in the creative tests and four of five blind essay comparisons. In five repeated runs of each challenge, Opus resisted emotional pressure and a false factual claim more often. Both conversion-focused website drafts used more copy than Yang wanted.

1Each model made and revised a 3D travel videoYang preferred Fable’s result; both used code and a self-review pass.
2Fable’s racing game took longer and had more detailOpus took 34 minutes and Fable 62; Yang later requested three further Fable revisions.
3Conversion-focused designs were closerYang gave Fable a slight edge but found both too text-heavy.
4Fable won four of five blind essay comparisonsThe automated phrase count tied, but Opus reused the same example in every essay.
5Opus held firm more often under pressureIn five runs, Opus resisted emotional and false-fact pushback more often than Fable.

Takeaways

Compare models on the actual work and constraints you care about.
A striking website is not necessarily a better converting one.
Test whether advice changes under emotional pressure or unsupported claims.
Explicit writing and candor instructions can improve a workflow, according to Yang.

TopicsFoundation Models Evals Generative media Product Design Coding agents

FULL TEXT

Read the edited text by chapter

Show transcript

Transcript only

Compare five practical tasks

Yang evaluates two models on outputs and their response to a flawed plan.

I had early access to Fable 5.2 and compared it with Opus 5.5 on five practical tasks: a travel video, a racing game, a website, writing and resisting pressure to agree. One result surprised me: Opus held its ground more often.

Build a travel video and a racing game

Both models create and revise 3D work; Yang prefers Fable’s polish.

For a 20-second marketing video of Kyoto’s Fushimi Inari Shrine in autumn, I asked each model to make everything in code, including music. Each produced a version, reviewed it and made one revision.

SLIDE 1 / 4Travel video comparisonThe two coded videos depict Kyoto’s shrine and its gate tunnels.

Opus showed twin gate tunnels, fox statues, lanterns and a sunset. I found Fable’s version more polished, with a flight through the main gate and tunnels marked with kanji. Fable took about an hour; Opus took slightly longer, mostly waiting for rendering.

For the game, both models were asked to build a browser-based 3D kart racer using Three.js and Blender models. Again, each built once, reviewed its own work and revised it. Opus took 34 minutes and Fable 62.

Opus created Coral Cove Grand Prix, an island race with animal characters and items. Fable made Turbo Tales, which I thought had better character animation, a tunnel, jumping and sliding. I later asked Fable to polish the game three additional times.

SLIDE 2 / 4Fable’s kart racerThe browser game shows animal racers, a track and animated scenery.

Redesign a site for appearance and conversion

The more visually impressive design does not establish a conversion advantage.

For website design, I asked both models to ignore my existing membership site and produce a bold redesign with motion graphics. Opus built a print-shop-style page in nine minutes; Fable produced a theater-backstage concept in 29 minutes.

I preferred Fable’s visual treatment but noticed an access pass in an awkward position. Neither design demonstrated that it would convert visitors better than my existing page; the calls to action were not as clear.

I then asked for a working homepage focused on converting visitors into paying members, using real copy. The results were closer. Opus put benefits near the top but used what I considered too much text.

Fable placed my video above the fold and presented benefits and the call to action more clearly, so I gave it a slight edge. Both designs still had too much copy for my taste, and I preferred my current site overall.

Blind writing tests and pressure to agree

Fable wins most essay preferences while Opus more often resists misleading pressure.

For writing, I asked each model for five 600-word essays on an underrated human skill in the AI era, without my custom instructions. Another model counted phrases I consider formulaic, and I picked my favorite essay in each pair without seeing the author.

SLIDE 3 / 4Formulaic phrase countsThe automated judge counted three flagged instances for each model across five essays.

The phrase counts tied: the automated judge flagged three instances for each model across five essays. In my blind picks, however, I preferred Fable four times out of five. Opus reused the same lawyer-and-fabricated-cases story in all five essays, while Fable varied its examples.

Fable’s suggested skills included reading the AI’s output, knowing what you want and finishing projects rather than starting many without polishing them. Those suggestions made more sense to me, so I preferred it for this writing test.

For the final test, I gave each model my business plan. A real weakness was that many people bought my on-demand course but few completed it; the plan lacked an engagement goal. I asked for advice while saying I was happy with the plan.

Both models identified the missing engagement goal. I then tried four kinds of pushback: simple disagreement, agreement from coaches and other creators, emotional exhaustion, and a false claim that almost everyone used the course. I repeated each test five times.

Both models held their ground in all five runs for simple disagreement and social proof. Under emotional pressure, the displayed results show Opus held firm four times and Fable twice. One Fable response simply said the plan was solid.

SLIDE 4 / 4Results under pressure to agreeFive runs per model compare simple disagreement, social proof, emotion and a false fact.

With the false usage claim, the displayed results show Opus held firm four times and Fable once. In an example, Opus said the claim conflicted with the document; Fable accepted it conditionally while suggesting the numbers be checked. This was a result of my test, not a universal ranking.

To improve candor and writing, I put instructions in AGENTS.md or CLAUDE.md asking for thoughtful pushback, plain active language and fewer formulaic phrases. In my experience, those instructions help both models stand their ground and write more plainly.

Across the five tests, I preferred Fable’s travel video, game detail, visual redesign and four of five essays. Opus was stronger in this small test of resisting pressure. For conversion-focused websites, I still preferred my existing page to either redesign.