muse spark 1.2 vs gpt 5.6 sol vs kimi k3 vs grok 4.5 – on two landing pages

our coding agents built two desktop landing pages from scratch, then had to open them in a real browser, find their own bugs and fix them before they were allowed to hand anything over

the setup: each model in its own vendor cli – @meta's muse code, @OpenAI's codex, @Kimi_Moonshot's kimi code, @spacexai's grok build – all pointed at one @openrouter key. identical brief, md5-checked, no hints

tasks:

1. lumen – a landing page for a money app. twelve sections, each needs its own entrance, its own interaction and its own idle motion, and no two sections may reuse the same interaction pattern

2. objkt – a headphones product page where every image is drawn in code. no photos, no external files, no libraries. one html file that works offline

the binding rule: two full self-qa passes in a live browser before shipping. find your own defects, fix them, report symptom, cause, what changed. shipping unverified is a fail

• total cost, both pages
#1 grok 4.5 – $2.48
#2 muse spark 1.2 – $4.70
#3 gpt 5.6 sol – $15.58
#4 kimi k3 – $19.19

• total build time, both pages
#1 grok 4.5 – 35m 50s
#2 muse spark 1.2 – 38m 07s
#3 gpt 5.6 sol – 55m 21s
#4 kimi k3 – 130m 23s

• total tokens
#1 grok 4.5 – 3,346,792
#2 gpt 5.6 sol – 15,490,610
#3 muse spark 1.2 – 16,420,166
#4 kimi k3 – 38,633,695

• lines shipped
#1 gpt 5.6 sol – 2,445
#2 muse spark 1.2 – 2,935
#3 kimi k3 – 4,991
#4 grok 4.5 – 6,439

• defects each model found and fixed
#1 gpt 5.6 sol – 22
#2 grok 4.5 – 18
#3 kimi k3 – 16
#4 muse spark 1.2 – 11

observations:

• the fastest single build in the whole benchmark was muse spark 1.2 – 14m 22s for the headphones page.

• it also did the one thing no other model did: when its own gallery test expected slide 1 and got slide 2, it worked out that the test was wrong and the page was right, because drag inertia carries further than one slide, and fixed the test!

• grok build did both pages for $2.48 and shipped more code than anyone – 1.9x cheaper than muse spark 1.2, 6.3x cheaper than gpt 5.6 sol, 7.7x cheaper than kimi k3

• grok 4.5 ran both builds at once and finished both inside twenty minutes – two landing pages in the clock most of this field spent on one.

• kimi k3 never compacts its context. on one task its spend per ten minutes ran 0.41, 1.33, 1.77, 2.51, 2.72. an agent that gets more expensive the longer it works is a budget problem before it is a quality problem

watch the full test via link