Tech • AI • Robotics • Game

ENFR

Claude Fable 5.2 Can Now Build Insane Games!

7/10
AIWorldofAINovember 14, 2026 at 10:55 AM8:01
Audio player

TL;DR

A neon racing game called Hyperloom was built largely through a single AI-driven workflow that combined Claude Fable 5.2 for planning, coding and testing with Higsfield for image, voice, 3D and trailer generation.

KEY POINTS

One-chat game production

Hyperloom was presented as a playable futuristic racing game created without manual coding. The workflow relied on Claude Fable 5.2, described as a model built for long-running tasks that can plan work, write code, test results and correct its own errors. The result was framed as a game production pipeline compressed into a single chat-based process.

AI split between code and media

The division of labor was clear: Claude handled reasoning, programming and validation, while Higsfield supplied creative assets through its MCP connector. The platform was described as linking more than 35 image, video and audio models into Claude, allowing one workflow to generate art, voice lines, 3D assets and promotional video alongside the codebase.

Initial concept art set the visual target

Development began with a single piece of key art for Hyperloom showing a sleek ship racing through a neon city at night with rain, wet reflections and rivals close behind. That image reportedly took about 90 seconds to generate at a cost of 2.75 credits. It then served as the visual blueprint for the rest of the project, defining ship shape, lighting and color palette.

Full racing game built in one extended run

The build request called for a complete 3D racing game with a city track, banked spiral, tunnel, jump, boost pads, drifting, five AI rivals and three laps. Audio was also generated in code rather than imported as files, including sound effects and the soundtrack. Before delivery, the system was instructed to run an automated driver through three laps to confirm the build worked.

Autonomous planning and testing

The coding model reportedly laid out a seven-step plan covering track design, city environment, ships, rivals, sound, controls, testing and publishing. It then worked unattended for about an hour, writing code, taking screenshots of the game, reviewing them and adjusting what looked wrong. The first pass returned with 50 out of 50 checks passing and no errors across keyboard, gamepad and touch input.

Playable first draft already included core features

The initial version of Hyperloom was described as fully playable and already included a mini-map, lap times and a boost meter. This first draft established the base mechanics before further changes were requested in plain language rather than with specialist game development terms.

Second round added voice, title screen and AI takeover

The next revision focused on feel and presentation: faster sensation, more aggressive rival behavior, an announcer countdown, a title screen using the key art and an AI autopilot if the player stayed idle. For the announcer, Higsfield generated voice lines including the race countdown. The idle takeover feature allowed the AI to continue racing on its own and, in one cited run, recover from last place to first.

From 2D art to 3D ship model

Converting the hero ship into a real in-game asset required an extra step because the rainy key art was too noisy for direct conversion. Clean views of the ship were generated first, then turned into a 3D model and rendered for verification. The updated game was then published online through Higsfield so the new files could be integrated and tested.

Expanded validation and more dynamic races

After the second round of upgrades, the automated test cycle expanded to 66 out of 66 checks completed successfully. Rival ships were also said to battle more actively, swapping positions 62 times in one race. Those changes were used to argue that AI iteration was not limited to static asset generation but could also tune gameplay behavior.

Trailer creation and pricing push

A final task asked the system to turn the key art into a short trailer using Seedance 2.5, with sound included. The broader pitch emphasized that work typically spread across concept art, programming, sound design, voice acting, 3D modeling and trailer editing could be condensed into one workflow. Higsfield was promoted with a free tier and paid plans, while the top Ultra plan was described as including all models and 3,000 credits per month, alongside a Black Friday discount of up to 60% off.

CONCLUSION

The demonstration highlighted how fast AI workflows are moving from isolated asset generation to end-to-end game production. The central question is no longer whether AI can help build a game, but how much of the traditional studio pipeline can be automated in a single toolchain.

Ask a question

More from AI