Claude Fable 5 roundup shows iOS ports, browser games, and 3D worlds

• minchoi's roundup framed the Fable 5 burst as 10 builds in under 99 hours, spanning native iOS ports, browser games, procedural 3D worlds, dungeon generation, and Blender assets.
• om_patel5's browser-racing breakdown supplied the most practical game-dev recipe: Sketchfab models, Poly Haven textures, Claude-compressed assets, Rapier physics, and a $20 Claude Pro plan.
• MengTo's workflow note described a web-design pipeline built from screen recording, video generation, and an unusually detailed prompt, while MengTo's full Koisei prompt specified three.js, GSAP ScrollTrigger, Lenis, WebGL shaders, and scroll-scrubbed video.
• LLMJunky's StarSwap run used Fable as an orchestrator over GPT 5.5 subagents and worktrees, ending with 40,000 lines of code and 96% of one 5-hour Pro session consumed.
• gregisenberg's PSA said the subscription window ended July 7 before pay-per-use credits, while AIandDesign's caveat said viral game demos still need human tuning for difficulty curves.

In minchoi's first example, Claude ported Command & Conquer to iPhone and iPad as native ARM64, not an emulator. bilawalsidhu's Ray-Bans build turned two pairs of Meta glasses into a 4D reconstruction pipeline. MengTo's Koisei prompt reads like a production brief for an interactive studio site, and Claude Fable 5 Prompt Library packaged reusable prompts for Claude Code workflows.

The game examples were strongest when they named the target surface and the constraint.

• minchoi's Command & Conquer example claimed a native ARM64 iPhone and iPad port with full RTS controls.
• minchoi's Subway Surfers example compressed the process to idea -> art -> 3D models -> code in 1 hour.
• minchoi's toy-to-game example said Claude turned a kid's toy into a mobile game, then added a volcano with one more prompt.
• According to om_patel5's racing-game post, the browser racer used downloaded cars, guardrails, streetlights, compressed textures, and Rapier physics so the car handled instead of sliding.
• minchoi's earlier games thread said people were building Claude Fable 5 games and some were making money with them.

That browser-racing clip is the cleanest creator takeaway: the model did not replace the asset pipeline, it squeezed it into a browser-friendly shape.

The 3D examples split into two camps: generated environments and lightweight asset tools.

• minchoi's ocean wildlife app paired AI videos with seamless transitions.
• minchoi's math-world example said Claude built models, animations, and effects from two prompts without imported assets.
• minchoi's FPS example described a procedural shooter mechanic with no assets, just prompts.
• minchoi's dungeon-generator example used Three.js for real-time dungeon generation.
• minchoi's dense-world example claimed a world larger than GTA 6 while still running at high FPS.
• minchoi's Blender-assets example pointed at game-ready Blender assets from simple prompts.

om_patel5's Topomaker post added a different 3D workflow: a browser modeling app with face coloring, timeline animation, and export to GLB, OBJ, GIF, or MP4. The stated target was web game developers and three.js projects, which makes it closer to a small asset workbench than a showcase demo.

MengTo treated the prompt as the build artifact.

His workflow had three parts:

• MengTo's reply said the detailed prompt came from a screen recording and a shorter prompt.
• MengTo's process post combined screen recording to prompt, video generation, scroll interaction specs, library choices, and section-by-section storytelling.
• MengTo's prompt-process reply said the work moved through assets, source references, and prompt details before iterating the final result.

The Koisei spec named the stack directly: vanilla HTML/CSS/JS or Vite, three.js, GSAP ScrollTrigger, Lenis, scroll-scrubbed video, custom shaders, and no UI frameworks, according to MengTo's full prompt.

The cost logic was explicit. MengTo's cost reply said one long prompt was preferable to 10 prompts that send an agent down a rabbit hole, and MengTo's brief-to-html reply put the sequence as brief -> prompt -> HTML -> iterations.

Fable also showed up as a planner above other agents.

• danshipper's button-color post joked that Fable spun up 100 agents to change one button color.
• LLMJunky's StarSwap workflow had Fable propose 10 improvements, choose 7, write independent plans into the repo, delegate implementation to GPT 5.5 High via Codex Exec, review and fix the outputs, then merge worktrees.
• thekitze's Benji rewrite paired Fable 5 with GPT 5.5 subagents for a TanStack Router and Convex rebuild.
• bas_fijneman's UGC Temple post described a complete MCP so any agent could use the project, with self-hosted and hosted options planned.
• According to shannholmberg's Ultracode explainer, Ultracode fans out a fleet of agents, has each take a slice, checks findings adversarially, and spends far more compute for a more thoroughly covered result.

The funniest demos were about output polish. The deeper workflow shift was delegation: Fable as reviewer, scheduler, and merge-conflict janitor.

The strongest non-game examples came from video and spatial capture.

bilawalsidhu's IronSight post fused footage from two pairs of Meta Ray-Bans into a 4D reconstruction. bilawalsidhu's build-time reply said the work started with queued experiments to test ideas and select primitives, then took a few hours the next day, and bilawalsidhu's hardware reply said the footage was all from the Meta glasses.

higgsfield_ai's Supercomputer post showed a community app that blocks 3D scenes, records a real cinema-camera move, and renders through Seedance 2.0. The 15-second clay pass preserved trajectory, timing, and framing, and higgsfield_ai linked the Supercomputer community app.

ai_artworkgen's character-performance post laid out a full character pipeline: original character in Midjourney, character sheet and scene expansion with OpenAI, performance in Seedance 2.0, and orchestration in Claude Fable 5. higgsfield_ai's comparison also put Fable 5 next to Sonnet 5 on agentic prompting for Seedance 2.0 outputs.

The hype arrived with meter anxiety attached.

• gregisenberg's PSA said Fable 5 was included through July 7, then moved to pay-per-use credits at $10/$50 per million tokens.
• gregisenberg's same PSA said it drew down usage roughly 2x faster than Opus and only ran up to 50% of the weekly limit during the window.
• shannholmberg's efficiency thread described Fable 5 as around 2x Opus in cost and said low effort could beat older models on high effort.
• shannholmberg's gotcha claimed sensitive requests could silently hand off to Opus 4.8 in the app, while the API exposes a signal.
• danshipper's fallback reply disputed a separate claim by saying it was the same model, but that it falls back to Opus 4.8 slightly more, so some benchmarks measure a mix.
• thekitze's token-burn post said Fable was eating tokens and burning money even when instructed to use subagents.
• bas_fijneman's reminder told people to use Fable while it was still available through a Claude subscription.

kaigani's Sayonara post captured the mood more neatly than any pricing table: people were already posting like a temporary collaborator was about to leave.

The counterexamples were useful because they named what the viral clips leave out.

LLMJunky's Fable result called one first-try website ugly, and LLMJunky's GPT 5.5 comparison said GPT 5.5 High did a much better first pass on the same comparison.

AIandDesign's game-design thread said Fable can produce something cool for a viral post, but it will not tune the game and difficulty curve into something enjoyable. The same thread said Radial Drift took over a month of tuning.

AIandDesign's one-shot reply doubted a polished result was truly one-shot unless many assets were already prepared, while AIandDesign's asset-prep follow-up said good prepared assets made a similar build easy with Opus in a few hours.

The last new wrinkle came from AIandDesign's Deadfall reply: after Opus failed at level design, he built a level-designer UI manually, then planned to ask Fable to generate levels automatically.