Andrew Warner and Jorge Sanchez watched a stack of Claude Fable 5.2 demos so builders do not have to judge a model by its prettiest clip alone. Their tour moves from Riley Brown's 4K animation to Alex Finn's agent planning, Peter Yang's website tests, Matt Wolfe's gadget files, a furniture-assembly rescue and Greg Isenberg's freelance draft. The recurring question is useful: what can you take into a real workflow today, and where does the human still have to check the work?
Fable 5.2 looks most promising here for ambiguous visual builds and substantial planning. Use it to make a first film, design a prototype or break a complex project into testable tasks. Do not treat a striking demo as evidence of conversion, a safe physical product or a paid freelance outcome. Andrew calls its visual ability "artificial general visual intelligence"; that is his description, not an official model classification.
Watch Andrew and Jorge's Full Discussion
The examples below are attributed to their original creators. The episode is a curated reaction and discussion, not a controlled benchmark or independent reproduction of every result.
1. Video and Animation: A Real Production Pipeline
Riley Brown's Fable film is the visual centerpiece. He starts with a broad creative brief, then has the model write Blender Python to construct a scene, place screen footage in a 3D display, animate lighting and cameras, and add audio. He reports rendering 1,329 4K frames with ten RunPod GPU instances in 52 minutes at roughly $6 in rendering charges. That is Riley's reported render cost, not a measured all-in cost: model tokens, sound, labor, storage and revisions also count.
Jorge's own practical use is narrower and easier to copy: app mascot animation and website motion. He says his Fable results are a large step up from what he produced with Opus 5.5 a few days earlier. For a first experiment, give Fable one real product asset and a short motion brief, render a ten-second sample, then inspect readability, brand consistency, frame pacing and mobile performance before scaling it into a full campaign. The later Opus-versus-Fable film comparison also shows why a beautiful sequence still needs a story and an ending.
2. Software Planning: Let the Expensive Model Define the Work
Alex Finn's segment demonstrates a useful division of labor: ask the strongest model to challenge the architecture, define milestones and write tasks with acceptance criteria; then hand the scoped implementation to less expensive models or developers. Fable pushes back on an overgrown agent team and proposes a simpler human decision queue. In the clip, the output becomes a detailed task board. It is evidence of a planning workflow, not proof that the whole agent harness was built.
For your own project, ask for a one-page plan, a dependency order, a list of irreversible decisions, and a testable "done" condition for every task. Make the executor attach tests and screenshots to each result. Anthropic's release guidance recommends Opus 5.5 as a starting point for most coding and knowledge work, moving up to Fable for ambitious, open-ended technical jobs. That makes model routing an economic choice, not a status symbol.
3. Website Design: Judge the Page by Its Job
Peter Yang's redesign test gives each model a visual direction, then a conversion-focused direction. Fable makes a striking 3D page, but Andrew and Jorge struggle to find the offer and primary action. The follow-up versions improve intent but become copy-heavy. Peter reports roughly nine minutes for an Opus pass and 29 for a Fable pass in this test; those are single-run observations, not a general speed ranking.
A useful review asks: can a first-time visitor state the offer, find the primary call to action and complete it on mobile? Check accessibility and load time too. Impeccable can help a coding agent critique visual quality, but its polish is not a substitute for real conversion data. Test the page with people who have not seen the prompt.
4. Physical Work: Design Files Are Not a Validated Device
Matt Wolfe's gadget experiment yields a parts list, circuit and firmware files, printable enclosure and an assembly guide. That is a substantial design package. The video does not show Matt fabricating, electrically testing or certifying this particular device. A real build still needs component checks, schematic review, fit tolerances, power and thermal testing, firmware validation, and relevant compliance work. Anthropic separately documents physical prototypes built with Diode Computers; those are a different project and do not validate every generated design.
How I AI's furniture example is more immediately actionable: ask the model to infer an assembly sequence from parts and photos, have it identify uncertain steps, then let a person carry out and correct the instructions. That is a strong assistant use case precisely because the physical world supplies feedback. Keep the person responsible for load-bearing, electrical or otherwise hazardous steps.
5. Freelance Work: A Draft Is Not Revenue
Greg Isenberg's prompt asks Fable to find a real freelance job and attempt it end to end. The example is an equity-analysis report listed at a $450 fixed price. The model produces draft work, but the episode does not demonstrate client approval, payment or a professional quality review. That distinction matters. For financial research, verify filings, dates, calculations and sources; have a qualified person sign off before delivery. Sell a defined, reviewed outcome rather than a promise that an agent can replace the analyst.
What We Would Try First
| Goal | Good first Fable task | Evidence needed before shipping |
|---|---|---|
| Marketing video | Storyboard and render a ten-second scene from owned assets. | Rights, story clarity, sound mix, render cost and mobile view. |
| Software build | Challenge the architecture and write acceptance-tested tasks. | Working code, tests, security review and task-by-task sign-off. |
| Website redesign | Create one bold and one conversion-led variant. | User comprehension, accessible CTA and actual conversion results. |
| Physical prototype | Draft CAD, bill of materials and an uncertainty list. | Engineer review, bench tests and, where relevant, certification. |
| Service offer | Produce a sample deliverable for an existing buyer brief. | Source checks, human approval and evidence of willingness to pay. |
The common pattern is to spend model intelligence where ambiguity is high, then narrow the job and verify the output in the world it must serve. That is more useful than asking whether Fable "beats" Opus in the abstract.
Watch the Six Original Demos
These are the creator videos Andrew and Jorge discuss. Open a clip to inspect the full source rather than relying on a short excerpt in the roundup.
Episode Chapters
| Time | Topic |
|---|---|
| 00:00 | Fable 5.2 and Andrew's visual-intelligence claim |
| 00:27 | Jorge Sanchez joins |
| 00:36 | Riley Brown's marketing video |
| 01:12 | The 4K result |
| 02:15 | Design and animation |
| 02:33 | Mascots and websites |
| 03:09 | Visualizing the workflow |
| 03:36 | Blender and Python |
| 04:21 | Music, effects and rendering |
| 04:30 | Rented GPU rendering |
| 05:24 | Pelican visual benchmark |
| 06:36 | Software projects |
| 06:54 | Plan before execution |
| 07:30 | Personal agent harness |
| 08:15 | Fewer agents |
| 09:09 | Chief-of-staff agent |
| 10:21 | Guided decisions |
| 10:39 | Notion task board |
| 11:24 | Tasks and acceptance criteria |
| 12:09 | Tool-platform segment |
| 13:03 | People search and integrations |
| 14:06 | Peter Yang's redesign |
| 14:24 | Opus versus Fable |
| 15:36 | Fable's 3D website |
| 16:48 | Beauty versus conversions |
| 17:24 | Conversion-focused redesign |
| 19:03 | Copy and page design |
| 20:06 | Beyond one prompt |
| 20:33 | Impeccable design review |
| 21:09 | Matt Wolfe's gadget |
| 22:03 | Firmware, CAD and parts |
| 23:15 | Printable product files |
| 23:51 | Assembly manual |
| 25:12 | Hardware possibilities |
| 25:39 | Furniture assembly |
| 27:09 | Human-in-the-loop work |
| 27:45 | Greg Isenberg's freelance idea |
| 28:03 | The real-job test |
| 29:06 | Draft deliverables |
| 30:27 | Marketing-film comparison |
| 31:03 | Opus versus Fable film |
| 32:15 | Reviewing the result |
| 33:36 | A story needs an ending |
| 34:21 | Using remaining tokens |
| 35:33 | Jorge's experience |
| 35:51 | Loggit finance app |
| 36:18 | More agent workflows |
Sources and Editorial Limits
This article uses the full Andrew Warner and Jorge Sanchez episode and the six linked original creator videos. Product status and model-choice guidance were checked against Anthropic's November 9 release; the Blender, RunPod and Impeccable links lead to the respective tool owners. Time, cost and quality comparisons in the videos are creator-reported demonstrations, not independently reproduced benchmarks. We have deliberately separated generated files from verified outcomes and visual quality from business results.