July 10, 2026
Ctrl+Alt+Delusion
GPT-5.6, Grok 4.5, Claude, and Muse Spark build the same 4 apps
AI app showdown gets messier as commenters roast the test, the prices, and the so-called winners
TLDR: A new AI showdown had 12 tools build the same apps five times each, and GPT-5.6 mostly came out looking strong while Grok impressed for the price. But the real battle was in the comments, where people argued the test was too clone-heavy, dunked on weak performers, and even fought over the price symbols.
The latest AI build battle tried to answer a simple question: if you ask a bunch of chatbot tools to make the same four mini-apps, who actually delivers? The creators went bigger this time with 12 models, five tries each, and a very public admission that this is not a lab-grade science fair. That honesty, weirdly, became one of the biggest talking points. One commenter basically said, stop apologizing — this is real investigation, just in a messy new field. In internet terms: the disclaimer itself started a debate.
And oh, the crowd had opinions. One camp rolled its eyes at the whole format, saying these tests are boxed in because the apps are all basically clones. Another camp loved the side-by-side chaos and started posting rival arenas and scoreboards, eager to compare the new GPT-5.6 trio. The nastiest little jab landed on GLM, with one commenter suggesting its flashy benchmark scores may not mean much in the real world after it went 0 for 5 on the maze. Ouch.
The funniest mini-drama? Someone noticed the pricing symbol and called out what looked like a cents-versus-dollars typo, which is exactly the kind of delightfully petty comment-section energy these posts attract. On the actual results, GPT-5.6 looked strong, Grok got praise as a cheap overachiever, Claude had an unexpectedly shaky outing in places, and Meta’s Muse Spark became the chaos gremlin of the bunch: broken most of the time, but weirdly brilliant when it worked. Classic comment-fuel.
Key Points
- •The article reran its coding-model build-off with 12 models, four app tasks, and five attempts per task after receiving reader feedback on the earlier version.
- •The updated benchmark added open-weights models GLM-5.2, Qwen 3.7 Plus, DeepSeek V4 Pro, and Kimi K2.6, served via Fireworks.
- •The article states that the comparison is not presented as objective or scientific and that all raw attempts are published so readers can inspect the outputs directly.
- •In the Doom-style raycaster maze task, GPT-5.6 Sol and GPT-5.6 Luna each scored 5/5 playable runs, Grok 4.5 also scored 5/5, and GLM-5.2 scored 0/5.
- •In the 3D Rubik’s Cube task excerpt, Claude Fable 5 scored 5/5 clean solves, GPT-5.5 scored 4/5, Grok 4.5 scored 3/5, and Claude Opus 4.8 scored 0/5.