GPT-6 Astra's First 48 Hours: Playable 4D Chess, 334-Part Tesla Model, WebGPU Physics
GPT-6 Astra has been publicly available for roughly 48 hours. In that window, a handful of developers published projects that function as informal ceiling tests for the model’s coding capability.
What Got Built
4D Chess — A playable implementation available at 4d-chess.pages.dev. The game extends standard chess to multiple boards (up to eight), with pieces that can move between parallel planes. Each piece gains additional valid moves relative to classical chess. The rules are internally consistent and the game runs in the browser. The developer described the process as a conversation with Astra rather than manual code-writing.
Tesla Model X, 334 Parts — An interactive 3D breakdown of a Tesla Model X with 334 individually modeled components. Each part is selectable and labeled. The level of part granularity — not just “engine” but individual brackets, clips, and housing elements — required the model to reason about mechanical assembly structure, not just render a car silhouette.
WebGPU Jelly Simulation — A real-time physics simulation using the WebGPU API. Soft-body physics at interactive frame rates is a domain that requires correct shader code, accurate material parameters, and proper GPU pipeline configuration. Building it in a short session with an LLM is a concrete measure of how much of that stack Astra can hold in working memory.
Game Boy Portfolio — A personal portfolio site rendered as a Game Boy interface, complete with period-accurate color palette and UI conventions.
Miniature 3D World — A procedurally generated small-scale 3D environment with apparent scene coherence.
What This Signals
None of these are production systems. What they measure is Astra’s effective coding floor — the complexity of project a non-specialist can produce in a short session without manual debugging.
The 4D chess result is the most diagnostic. Implementing a novel ruleset in a working browser game requires: consistent rule encoding, state management across multiple boards, valid move generation for each piece type under extended rules, and UI that reflects the extended game state legibly. That Astra produced something playable rather than a visual mockup with broken game logic is a data point about its code coherence at project scale.
The 334-part Tesla model tests a different dimension: hierarchical reasoning about physical structure. LLMs have historically struggled with “part is a part of a part” decomposition at granular scales.
Arena.ai added Astra to the Code Arena WebDev leaderboard on September 5. Formal benchmark results will confirm or complicate the picture from these demos. Until then, the 48-hour output is the most accessible signal available.