AI Model Reviews

Fable 5.2 Demos: What Is Actually Useful?

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?

The useful answer

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

GoalGood first Fable taskEvidence needed before shipping
Marketing videoStoryboard and render a ten-second scene from owned assets.Rights, story clarity, sound mix, render cost and mobile view.
Software buildChallenge the architecture and write acceptance-tested tasks.Working code, tests, security review and task-by-task sign-off.
Website redesignCreate one bold and one conversion-led variant.User comprehension, accessible CTA and actual conversion results.
Physical prototypeDraft CAD, bill of materials and an uncertainty list.Engineer review, bench tests and, where relevant, certification.
Service offerProduce 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.

Riley Brown. A 4K marketing film assembled with Blender code and rented rendering. Inspect the actual workflow and costs, not only the final clip. Watch on YouTube.
Alex Finn. A strong model creates a personal-agent plan and detailed tasks. The segment shows planning, not a finished agent platform. Watch on YouTube.
Peter Yang. A direct Fable-versus-Opus comparison across five tasks, including visually bold and conversion-focused website redesigns. Watch on YouTube.
Matt Wolfe. A proposed gadget, firmware, printable parts and assembly documentation; physical construction is a separate step. Watch on YouTube.
How I AI. A furniture assembly problem where a person uses the model's instructions to complete the job. Watch on YouTube.
Greg Isenberg. A set of real-work prompts, including a freelance-report draft. No client acceptance or payment is shown. Watch on YouTube.

Episode Chapters

TimeTopic
00:00Fable 5.2 and Andrew's visual-intelligence claim
00:27Jorge Sanchez joins
00:36Riley Brown's marketing video
01:12The 4K result
02:15Design and animation
02:33Mascots and websites
03:09Visualizing the workflow
03:36Blender and Python
04:21Music, effects and rendering
04:30Rented GPU rendering
05:24Pelican visual benchmark
06:36Software projects
06:54Plan before execution
07:30Personal agent harness
08:15Fewer agents
09:09Chief-of-staff agent
10:21Guided decisions
10:39Notion task board
11:24Tasks and acceptance criteria
12:09Tool-platform segment
13:03People search and integrations
14:06Peter Yang's redesign
14:24Opus versus Fable
15:36Fable's 3D website
16:48Beauty versus conversions
17:24Conversion-focused redesign
19:03Copy and page design
20:06Beyond one prompt
20:33Impeccable design review
21:09Matt Wolfe's gadget
22:03Firmware, CAD and parts
23:15Printable product files
23:51Assembly manual
25:12Hardware possibilities
25:39Furniture assembly
27:09Human-in-the-loop work
27:45Greg Isenberg's freelance idea
28:03The real-job test
29:06Draft deliverables
30:27Marketing-film comparison
31:03Opus versus Fable film
32:15Reviewing the result
33:36A story needs an ending
34:21Using remaining tokens
35:33Jorge's experience
35:51Loggit finance app
36:18More 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.

Common questions

Is artificial general visual intelligence an official Fable 5.2 classification?
No. AGVI is Andrew Warner's phrase for the visual capabilities he sees in these demos, not an Anthropic product category or an established benchmark.
Did Riley Brown make a 4K film with one prompt for six dollars?
He describes a single starting brief, then a workflow involving Blender code, a screen recording, music and sound effects, and rented GPU rendering. The roughly $6 figure is his reported render cost, not the total cost of model usage, labor or production.
Did the demos prove that Fable 5.2 can replace designers, engineers or financial analysts?
No. The episode shows promising outputs and individual tests. It does not establish repeatable conversion lifts, a validated manufactured device, or a paid and accepted freelance deliverable.
Share
X LinkedIn Reddit
Build Yours

Want a system
like this one?

Book a free 30-minute call. We map your situation, identify the highest-impact automation, and figure out if we are a fit.

Book Free 30-min Call