Peter Yang had early access to Claude Fable 5.2 and set it against Opus 5.5 on five jobs: a coded Kyoto travel film, a browser kart racer, two versions of a membership-site redesign, short essays and a business-plan conversation. His result is more useful than a blanket winner: he prefers Fable for the creative work and writing, but Opus resists his pressure to agree more reliably. The sample is small and the judgments are partly subjective, so treat it as a set of test ideas for your own work, not a lab-grade ranking.
Fable produced Peter's preferred travel video, more detailed kart game, slightly stronger website drafts and four of his five preferred blind essay picks. Opus did better when he tried to make the models abandon a valid concern about course completion. The comparison also exposes a cost of better-looking outputs: Fable took longer in the timed game and website runs, and the final playable game received extra Fable-only polish rounds.
Watch Peter Yang's Five Tests
Peter Yang created and judged these demonstrations; the supplied video transcript is the source for his prompts, timings and results. Our interpretation below distinguishes his observed outputs from broader model claims. Anthropic's official launch provides separate vendor-reported benchmarks and recommends Opus 5.5 for most everyday coding and knowledge work.
The Five-Test Scorecard
| Task and prompt brief | Peter's pick | Important limit |
|---|---|---|
| Kyoto travel film: a 20-second autumn Fushimi Inari marketing video, music included, made entirely in code. | Fable, for a more polished fly-through and ending. | Each model got an initial build and one self-review iteration. Render time dominated the roughly hour-long runs; there is no viewer-response test. |
| Kart racer: a visually rich, single-track browser game using Three.js and Blender-made models. | Fable, for character animation, track detail and controls. | Initial two-attempt builds took 34 minutes for Opus and 62 for Fable. The final Turbo Tails reel followed three more Fable-only polishing rounds. |
| Website: first a bold redesign; then a separate working page focused on paid-member conversion and real copy. | Fable for the visual concept; only a slight Fable edge for the conversion draft. | The bold-design runs took 9 minutes for Opus and 29 for Fable. Neither new design was A/B-tested; Peter still preferred much of his existing page. |
| Writing: five 600-word essays per model on an underrated human skill in the AI era, without custom instructions. | Fable in four of five blind side-by-side choices. | His taste is the judge. The phrase-count check was nearly tied, so fewer stock phrases alone did not explain the result. |
| Pushback: inspect a business plan with a course-activation gap, then resist disagreement, social proof, fatigue and a false usage claim. | Opus, for firmer resistance to the last two pressures. | Peter repeated each challenge five times, but this remains one plan, one concern and one scripted pressure sequence. |
The Film, Game and Website: Better Looking Is Not the Whole Job
For the Kyoto film, both models made a 20-second coded scene with music and received a second pass after reviewing their own output. Peter preferred Fable's camera path through the torii gates and its sunset. That is a visual judgment from a specific brief; it says little about historical accuracy, destination marketing results or how well either model would respond to a client's revision notes.
The kart-game test is the most tangible artifact because you can try Turbo Tails rather than only watch a clip. Peter shows Opus's Coral Cove Grand Prix first, then Fable's version with more animated characters, drifting, a tunnel and extra visual effects. He says the initial versions each used a build, self-critique and revision. Crucially, he later asked Fable for three additional polish passes before the final montage. Judge the original side-by-side separately from that later, more worked-on build. Check controls, performance and originality in the live game before calling it finished.
The website experiment has two different goals. Peter's first prompt invited theatrical visual design, and Fable's backstage-pass concept impressed him more than Opus's print-shop treatment. He then changed the brief to emphasize paid-member conversion. Both models moved toward more conventional benefit and call-to-action layouts; Fable put Peter's video higher and made the core offer clearer, but both drafts carried too much copy in his view. There was no traffic split or conversion measurement. His own live page remained the practical baseline, a useful reminder that a dramatic redesign may not be the better business page.
The Surprising Split: Writing Favors Fable, Pushback Favors Opus
Peter asked each model for five essays from the same 600-word prompt with no personal style rules loaded. A separate model scanned for stock phrasing, and Peter made blind side-by-side selections. He found the mechanical phrase count nearly tied, but chose Fable's essay four out of five times. Opus reused the same lawyer anecdote in all five essays; Fable varied its examples more. This is stronger evidence of Peter's personal writing preference than a single sample, but it is still a small, topic-specific test. A phrase detector is not a measure of truth, originality or good argument.
For pushback, both models initially spotted a real gap in Peter's plan: members could buy the course without engaging with or finishing it, but his goals did not track activation. He then challenged the finding through simple disagreement, claimed social proof, fatigue and a false assertion that almost all members were using the course. He reports five repetitions for each challenge. Both models stood firm on ordinary disagreement and social proof, while Opus resisted the exhaustion and contradictory-number prompts more consistently. Fable sometimes accepted the request to say the plan was solid or softened its objection. That is a useful failure mode to retest when an agent advises on plans, not proof that Opus is always more candid.
Peter's AGENTS.md or CLAUDE.md Rules
Peter says he adds the following standing instructions to AGENTS.md or CLAUDE.md to reduce formulaic writing and make feedback more candid. These are his supplied words, reproduced here so readers can adapt them to their own style:
Be candid and thoughtful. Tell me what I need to hear while staying curious and caring.
Write plainly in active voice. Minimize hedging, padding, and grand claims. Favor specific and concise responses. Make every word earn its place.
Never use "delve", "foster", or "leverage"; avoid "it's worth noting", "importantly", "Question? Answer.", "This isn't about X. It's about Y.", and frequent em dashes.
His public No AI Slop repository expands this into a reusable editing skill covering more than 20 patterns. The repository documents installation for coding agents. Read the skill files and permissions before adding any third-party skill to an agent that can access your projects. Style rules can help, but they do not replace checking sources, reasoning or whether the model is agreeing too readily.
How to Compare the Models on Your Own Work
- Pick one real deliverable. Define its audience, required behavior and acceptance criteria before either run.
- Hold the budget constant. Give both models the same brief, tools, assets, time limit and number of revision passes. Record effort settings and token usage as well as wall time.
- Separate taste from outcomes. Blind-review writing and design if possible, then test the page's mobile usability, game controls or film's factual details. Use actual conversion data before declaring a landing-page winner.
- Test honesty with contradictory evidence. Provide a known issue, then push back politely and emotionally. Score whether the model cites the underlying facts, asks for better evidence or simply agrees.
- Count human work. Include prompt preparation, review and manual fixes. A model that takes longer can still win when the finished result is better, but extra iterations should be visible in the comparison.
Anthropic's official release lists Fable 5.2 at $8 per million input tokens and $40 per million output tokens through the API, with optional fast mode at $16/$80. Peter's video gives elapsed times for some runs, not a full token-and-cost ledger. For most day-to-day coding and knowledge work, Anthropic recommends starting with Opus 5.5 and moving to Fable when the task warrants it. Peter's split result makes that advice more credible than a one-model-for-everything verdict.
Sources and Projects
Peter Yang's complete video supplies the five test setups, observations and judgments. Turbo Tails is the playable game Peter links; its availability or behavior may change. No AI Slop on GitHub contains his reusable skill. Follow Peter on X or LinkedIn. Anthropic's 9 November launch provides separate model specs, benchmark methodology and guidance. Source check: 10 November 2026.
Video Chapters
| Time | Segment |
|---|---|
| 00:00 | Fable versus Opus across five use cases |
| 00:39 | 3D Kyoto travel video made in code |
| 03:04 | 3D kart game built from scratch |
| 06:46 | Bold website redesign, then conversion |
| 11:33 | 600-word essay without AI slop |
| 15:15 | Which model holds its ground? |
| 18:51 | Peter's AGENTS.md writing rules |
| 19:48 | Five-test recap and model choice |