RocketLeagueBench critics question GPT-6 Astra scores versus visible game quality

• RocketLeagueBench uses a familiar game as a visual-replication target. LLMJunky says in LLMJunky's benchmark rationale that Rocket League gives each run a known reference for direction-following and 3D reproduction.
• Astra received an unusually prescriptive build brief, as LLMJunky's full prompt calls for Three.js, Rapier physics, a complete gameplay loop, visual polish, and browser verification.
• The demo's physics remain self-assessed: LLMJunky's physics caveat flags a ball that can feel too bouncy, while LLMJunky's multiplayer clarification says the build is currently local-only.
• The reported spend fell from about $100 to closer to $25 when LLMJunky's cost correction accounted for a fraction of a subscription week and other work.

LLMJunky's full prompt requires the actual game to appear before any landing page, then asks the model to test the browser build and capture proof that the canvas is not blank. Three.js's WebGLRenderer docs describe its WebGL 2 scene rendering, while Rapier's JavaScript guide describes the rigid-body contacts and forces the brief delegates to the physics engine.

LLMJunky chose Rocket League because its arena, car handling, and visual language make the output easy to compare with an existing target. The creator says the test is deliberately simple because earlier models could not reliably handle more complexity, and says the benchmark may now be saturated enough to warrant harder subjects.

The creator says the same test had already been run more than a dozen times, with earlier results never actually resembling Rocket League, in a longer explanation. That makes the clip a useful progress artifact, but it measures a narrow combination of reference matching, browser implementation, and visual polish.

LLMJunky posted the specification after a reader asked about the prompting in a reply asking how Astra was prompted. Its structure is the practical artifact here:

1. Stack and architecture: Three.js for rendering, preferably Rapier.js for physics, TypeScript, Vite, and separate rendering, physics, controls, state, UI, and assets.
2. Playable loop: controllable car, ball and arena collisions, scoring and resets, boost pads, jumping, flips, air control, timers, camera modes, and an optional bot.
3. Presentation: stadium lighting, colored goals, trails, bloom if performance allows, particles, a HUD, and laptop-screen resizing.
4. Verification: run the dev server, play it in-browser, check movement, collisions, scoring, camera behavior, console errors, and capture a nonblank canvas.

LLMJunky identifies the model setting as Astra xhigh in a configuration reply. The last paragraphs add a repeated iteration loop aimed at matching Rocket League's handling, car, arena, and UI rather than accepting a first usable pass.

LLMJunky says the prototype can air dribble and flip reset in a mechanics reply, but also names ball restitution, tackles, and the car's landing response as remaining quirks. The creator described the implementation as browser TypeScript in a technical reply.

The timeline is similarly narrow: LLMJunky estimated four or five hours of work in a timing reply. Multiplayer was absent, so the reported latency was only that of the local machine, according to LLMJunky's multiplayer clarification; the creator called netcode the hardest piece.

LLMJunky says Meta Muse Spark 1.3 Max and Astra received the same prompt in a comparison reply. A separate reply says Muse took a couple of iterations while Astra's base output looked strong from the start in a follow-up.

No numeric RocketLeagueBench score, scoring rubric, or run log appears in the supplied thread. The creator also says several earlier models performed better than the displayed Muse result in past runs in another follow-up, without linking those results.

MindStudio's separate analysis made a similar claim about a different test, arguing that Muse Spark 1.3's reported 75.4% DeepSWE score did not predict its project output. It does not independently test RocketLeagueBench.

LLMJunky first put the work at around $100, then corrected it to closer to $25. The creator says the accounting came from using about 53% of one week on a $200 plan while doing other work, in a billing reply, so it does not yield a clean per-run price.

In response to a request to open-source the game, LLMJunky said it would not be published in a publication reply. The full brief asks for near-identical Rocket League presentation, and the creator later said a distributable car-soccer game would need its own car and field rather than exact Rocket League assets in an IP constraint reply.

LLMJunky had already asserted that no open-source Rocket League code existed for the model to draw from, saying it used vision and wrote TypeScript in an earlier reply. During a dispute with another car-soccer project, LLMJunky said the variants were different in a response, then described this build as Three.js and Rapier in TypeScript versus the other project's Unity stack in a follow-up.

The audit shown in that screenshot reports no identified reuse in the current codebase or Git history, while explicitly saying it was not a line-by-line similarity audit.

LLMJunky says the benchmark experiments also reached beyond the browser stack: they tried Unreal and Godot, did not try Unity, and were surprised by how well Godot and its MCP integration worked in LLMJunky's engine tests.