
Tech • AI • Robotics • Game
Abacus AI says its GPT-6 Astra workflow can design, build, test, optimize and deploy original browser-based 3D games, demonstrated through a 5v5 MOBA, a dark fantasy action RPG, and a formula-style racing game.
Astra begins by generating concept images for characters, environments, lighting and camera framing, then uses those images as visual targets during development. It writes Blender scripts to create original 3D assets, exports them into a browser game stack using Three.js and Rapier, and adds materials, shaders, post-processing and physics. It also launches builds, captures screenshots, compares results against the original visual brief, and iterates when scenes look wrong.
The workflow includes automated tests for gameplay logic, physics, track geometry and other systems, with projects running dozens of checks during development. When performance drops, it can optimize with instancing, texture atlases, LOD systems, WebP compression and asset deduplication. It also pulls openly licensed textures through Poly Haven when realistic surfaces such as asphalt or forest floors are needed.
The largest demo, Emberfall, is a browser-based 5v5 MOBA with two bases, three lanes, a jungle, turrets, inhibitors, minion waves, neutral monsters, a boss, bushes with vision mechanics, recall, respawn, gold, XP, an item shop and a full HUD. It includes five original champions across tank, fighter, mage, marksman and assassin roles, each with a passive, three abilities and an ultimate mapped to Q, W, E and R. The playable build features champion select, enemy bots, damage numbers, kill announcements, a shop, a scoreboard and a minimap.
Development changes went beyond broad layout. Astra adjusted the mage’s shoulder proportions after identifying them as too wide, redesigned vegetation after recognizing that roughly 900 trees at about 5,000 triangles each would hurt browser performance, and replaced that approach with lighter instanced assets. It later improved champion armor, towers, ruins, water animation, foliage and color grading, and generated painted portraits plus 20 ability icons.
The second project, Eye of the Fallen God, shifts focus from systemic complexity to atmosphere and scale. Set in an ancient prison above an abyss, it centers on a giant stone god named Akara, built directly into the architecture above a ritual arena. The player controls the last descendant of the bloodline that imprisoned it, using full third-person combat with lock-on targeting, soul energy, spectral dash and soul cleave.
Before building, the system generated 10 concept shots covering major reveals, boss encounters and collapse sequences. Blender then produced the arena floor, pillars, arches, glyphs and the carved stone face as real geometry, supported by PBR materials and a seamless panoramic sky with moons and nebulae. Bosses include Varos, a melee fighter, Sarah, who attacks from platforms with spectral arrows, and Torin, a chained colossus built around a different encounter structure.
The system made highly specific corrections during iteration. It changed oversized pillar capitals, traced visual flicker to co-planar triangles in geometry files, rebuilt the sky when the world felt too flat, and fixed a one-frame cape animation jump by adjusting animation blending. One standout sequence starts in a dark corridor, then opens into the arena as the camera pulls back and tilts up toward Akara’s face before returning control to the player.
The third demo, Extreme Formula: Celestial Alps GP, centers on an original car named the Nova Celestus 07. Its design includes a low open-wheel chassis, slick tires, multi-element wings, halo protection, exposed push-rod suspension, a diffuser and aurora-style teal, magenta and violet livery. The circuit is a 1.27-kilometer closed loop with elevation change, forests, cliffs, a waterfall bridge, a village section and a tunnel.
The car is not a static prop: wheels rotate with road speed, the front wheels steer, and each wheel has independent suspension travel. The final game includes four camera modes—chase, close chase, cockpit and front wing—plus lap timing and best-time storage. After deployment, the system traced reversed steering inputs to the yaw sign, flipped it, rebuilt the project and redeployed the live version.
Abacus AI positions the system as a persistent production environment rather than a one-click AAA game generator. Two model roles split responsibilities between visual planning and code generation, while the hosted environment keeps project state, servers and deployment active between sessions. The company also acknowledges limits, noting these are not finished commercial games and that the system can flag when requests exceed what a single-session workflow can realistically deliver.
The demos suggest that AI-assisted game creation is moving beyond toy projects into iterative 3D production, testing and deployment. The main significance is not replacement of major studios, but the emergence of a faster prototyping pipeline that can build and refine surprisingly complex playable worlds.
Explain this