Same task, every model. What does the extra money buy?
Each prompt went to 16 models, asked 3 times each (a few runs are missing, and are counted as not run). You can read what was asked, but not how it was asked, so the prompts can't be trained on. Nothing is scored. You get the work, the price and your own eye.
Models listed cheapest first·How it works →
SVG
Draw a lighthouse on a rocky outcrop during a storm, as a single self-contained SVG
16 models · up to 3 runs each · median $0.0015 to $0.288 per answer (186×) · 5 not run (our error) · 1 hit our limit
None of the 42 finished drawings uses an SVG filter. Five models drew with gradients in at least one answer.
186×Gemini 3.1 Pro's median cost per answer, divided by Qwen3 235B A22B's
SVG
Produce a six-icon weather set on a shared grid, in one SVG
16 models · up to 3 runs each · median $0.0012 to $0.197 per answer (160×) · 5 not run (our error) · 3 hit our limits
160×Claude Opus 5's median cost per answer, divided by Qwen3 235B A22B's
HTML page
Build a complete three-tier pricing page as one self-contained HTML document
16 models · up to 3 runs each · median $0.0014 to $0.219 per answer (151×) · 2 not run (our error)
None of the 46 finished pages loads anything from another address. Five models labelled one tier “Most popular” in at least one answer.
151×Claude Opus 5's median cost per answer, divided by Qwen3 235B A22B's
Design component
Design a dense data table with sorting affordances and a selected row state
16 models · up to 3 runs each · median $0.0019 to $0.258 per answer (139×) · 12 not run (our error)
Of the 15 models with a finished table, 7 mark the sorted column with aria-sort in at least one answer, and 6 mark the selected row with aria-selected.
139×Claude Fable 5's median cost per answer, divided by Qwen3 235B A22B's
Joke
Write an original dad joke about programming
16 models · 3 runs each · median $0.000033 to $0.051 per answer (1,551×) · 5 hit our limits
The prompt asks for an original joke. Three models told a version of the Java/C# glasses joke, and Claude Sonnet 5 told the same dark-mode joke on all three runs.
1,551×Kimi K3's median cost per answer, divided by Qwen3 235B A22B's
Short text
Announce a fictional product in exactly one paragraph, without marketing clichés
16 models · 3 runs each · median $0.00012 to $0.039 per answer (334×) · 1 not run (our error) · 2 hit our limits
13 of 16 models used “Dawn” in at least one answer, often in the lamp’s name: Dawnlight, Dawnwell, DawnGlow. All 43 finished answers are one paragraph of 62 to 89 words.
Introducing the DawnGlow Desk Lamp, designed to mirror the natural sunrise with a gentle gradient of light that builds over thirty minutes, easing you out of sleep. Its warm, adjustable glow begins softly, increasing in intensity to help regulate your circadian rhythm, making mornings feel less abrupt. Compact and modern, it fits neatly on any nightstand or desk, doubling as focused task lighting by day. The DawnGlow blends form and function, turning wake-up time into a quieter, more natural transition.
Meet Dawnlight, a desk lamp that ends the alarm clock's tyranny. Thirty minutes before you need to be awake, it begins to glow—faint as first light—then builds gradually until your room feels like early morning, even in December. No sound, no jolt, no groping for a snooze button. Just your eyes opening on their own, the way they were meant to. Dawnlight ships this fall in brass, matte black, and birch. Preorders open Monday.
334×Claude Fable 5's median cost per answer, divided by Qwen3 235B A22B's