A recent evaluation pitted twelve AI models against four distinct software development tasks, including a Doom-style raycaster, a 3D Rubik's Cube solver, a functional calculator, and Conway's Game of Life. The test aimed to assess model performance across varied complexities, from basic logic to graphical rendering and animation.

Models included GPT-5.6 (in Sol, Terra, Luna variants), Grok 4.5, Claude Opus 4.8 and Fable 5, Meta's Muse Spark 1.1, and several open-weight models like Qwen 3.7 Plus and Deep Seek V4 Pro served via Fireworks. Each model received five attempts per task, with evaluators noting success rates, best builds, costs, and average times. Notably, GPT-5.6 Sol excelled in the raycaster and calculator tasks, while Claude Fable 5 achieved a perfect score on the Rubik's Cube. Open-weight models demonstrated competitive performance, particularly in simpler tasks like the Game of Life.

The results suggest a narrowing gap in capabilities, even between proprietary and open-weight models, across diverse application types. While GPT-5.6 generally led, specific tasks saw strong outputs from other models like Claude and even surprising performance from Meta's Muse Spark 1.1 on its successful runs. The evaluation emphasizes that model suitability remains task-dependent, with cost-effectiveness also a significant factor, as seen with Grok 4.5's strong showing for its price.