02 October 2026
Qwen 3.8 Flash Next
Sushi 1.1.0 · Pi coding agent
Local run
A three-inning game build with recorded headless game checks. The run sheet notes acceptable visuals and inconsistent baseball rule behaviour.
Workload · Agentic coding · 3D baseball · Revision 1
Hardball 3D
A one-shot coding challenge to build a complete, playable 3D arcade baseball game. The prompt asks for at least one full inning, with the player batting in the top half and pitching in the bottom half.
3 recorded attempts
The builds, side by side
04 October 2026
Qwen 3.8 27B
TensorFold 0.6.4 · Pi coding agent
Local run
Pi's final session report records gameplay checks for both batting sides, scoring, inning progression and restart, with no console or page errors reported during multi-minute autoplays.
03 October 2026
GPT-6 Luna
OpenAI Codex · xhigh reasoning · one-shot
Cloud baseline
A one-inning game build with a successful Vite production build recorded. No independent gameplay test report was retained.
From the run records
Measured side by side
Each attempt used its own engine, harness version and settings, and some sessions include waiting or outages. Hover a value for its exact definition, and read the run sheets before drawing conclusions. Cloud wall time includes provider queueing; local decode speed is per request on this machine.
Task requirements & full brief
Each attempt starts from a project scaffold and must deliver the full game loop: pitching, batting and timing, ball movement, fielding, base running, outs and scoring, inning changes, visible feedback, and restart without reloading. The game should use procedural Three.js geometry and work in a modern desktop browser.
- Control the swing and make timing affect contact.
- Show AI fielders moving to retrieve and return the ball.
- Track runners, occupied bases, outs, and score.
- Complete both halves of an inning and show the result.
- Keep the game restartable from the browser.
- Avoid external engines, game assets, and backend services.
How to read this comparison
Compare the generated builds and the checks recorded for each attempt. There is no normalised quality score, and the cloud task’s cumulative token totals and end-to-end rate include repeated context, reasoning, tool calls and waiting — they are not comparable with local per-request decode or prefill rates. The method page sets out which figures can and cannot be set against each other.