Gemini 3.1 Pro thinking: medium · multimodal
branch feature/inline-svgs-with-gemini-3.5-flash · 4 real prompts × 4 configs (3 models; GLM-5.2 in two OpenRouter routes) · cells are the live HTML (zoom freely, stays crisp) · open↗ = full size
| prompt | Gemini 3.1 Prothinking: medium · multimodal | Gemini 3.5 Flashthinking: medium · multimodal | GLM-5.2reasoning: medium · Z.AI · text-only | GLM-5.2reasoning: medium · nitro (throughput) · text-only |
|---|---|---|---|---|
| Garden Soiree | $0.170100s · 6 svg | $0.080646s · 5 svg | $0.0564127s · 13 svg | $0.032043s · 10 svg |
| Birthday Bash | $0.052425s · 8 svg | $0.10356s · 4 svg | $0.114323s · 9 svg | $0.024734s · 10 svg |
| The Crusty Loaf | $0.12981s · 10 svg | $0.12164s · 4 svg | $0.037686s · 9 svg | $0.025341s · 9 svg |
| Bright Smile | $0.13077s · 6 svg | $0.077046s · 8 svg | $0.0577161s · 10 svg | $0.026842s · 11 svg |
| TOTAL / column | $0.481tot · avg 71s · 30 svg | $0.381tot · avg 53s · 21 svg | $0.266tot · avg 174s · 41 svg | $0.109tot · avg 40s · 40 svg |
Prompts are the exact ones from the product's own dev runs. Each cell shows latency · cost · inline-svg count · output tokens · image kind. All runs: product flyer, A5 Portrait, temp 0.3, font web. Gemini = thinking_level medium; GLM-5.2 = OpenRouter reasoning medium, capped 32768 out, run two ways: Z.AI (first-party endpoint, pinned) and nitro (OpenRouter throughput route → fastest provider). Same model & prompt — only the upstream provider differs. GLM cost: Z.AI @ $1.40/M in · $4.40/M out; nitro @ ~$0.95/M in · $3.00/M out (throughput route, typically DeepInfra; approximate). Gemini cost via genai-prices.
On where the images come from: raster photos are real user-supplied cut-out assets (birthday-cake.png, 02_bakery.png) embedded inline as base64 — exactly as the live product does. The two Gemini models are multimodal and are shown the asset (-img); they place it and the product embeds it. GLM-5.2 is text-only — it never sees the asset, so on the image prompts it can only emit a slot (filled afterward with the same provided asset) or draw SVG. Garden & dental had no asset, so all models there are pure HTML/CSS+SVG (dental's QR is a locally-generated PNG). Nothing here is an invented/AI-generated image.