Harbor Cat Courier (2D pixel-art game)
We wrote these prompts and ran each one once, identically, on Gemini 4 Hydrogen, Gemini 3.8 Flash, and Gemma 4 31B through the Gemini API free tier on 14 November 2026 (America/Denver). Hydrogen and 3.8 Flash used thinking_level=MEDIUM; Gemma 4 has no thinking level. We did not edit the models’ code. Playable frames show the raw HTML or SVG (Three.js CDN URLs were rewritten to a local copy so the iframe does not call the public internet). Dollar figures are equivalent paid-tier list prices — labelled “equivalent paid cost; ran on free tier.” A 429 or quota error stops that model for the rest of the batch; leftover tests wait until the free-tier quota resets at midnight Pacific. Free-tier prompts may be used by Google to improve products. This is not a benchmark: one run per prompt, our tasks, our settings.
The prompt (copyable)
Write a complete playable single-file HTML game. No build step, no frameworks, and no external images, fonts, or scripts. Title it Harbor Cat Courier. The player is a cat on a bicycle delivering steamed buns: arrow keys or WASD move left and right and jump, catch falling buns, and dodge seagulls, in a 60-second score attack. Include a start screen and a game-over screen with a restart control. Draw every sprite with canvas pixel rectangles (no emoji as the only art). The file must run by opening it in a browser. Return only the HTML document, starting with <!DOCTYPE html>.
Side-by-side playable results
Each frame is sandboxed (scripts allowed, no same-origin). External network is blocked by CSP, except our local Three.js copy where needed.
Gemini 4 Hydrogen
ran54.8s · 11,409 tokens · $0.0846 equiv. paid cost; ran on free tier
maxOutputTokens=24576 · thinking_level=MEDIUM
A complete start screen shows score, best, a 60-second timer, a control legend, and a Start Delivery button. After start, a pixel-art cat on a bicycle rides a pier with a lighthouse and water, falling buns, and a countdown timer.
Gemini 3.8 Flash
errorThe API returned an error, including after one extra attempt. No output this session.
No output this session. The Gemini 3.8 Flash API returned HTTP 503 (“high demand”) or dropped the connection. We tried once more after the first failure and did not retry in a loop. This was not a quota / 429 error.
Gemma 4 31B
ran1m 51s · 3,930 tokens · $0.0000 equiv. paid cost; ran on free tier
maxOutputTokens=24576 · thinking_level=n/a
The start screen is titled HARBOR CAT COURIER and lists WASD or arrows. Two dark rectangles overlap the “PRESS ANY KEY TO START” line; they match obstacle blocks used in play. After a key press, a 60-second game starts: a small orange cat-like figure with two wheels, a falling bun, gray blocks, a score, and a countdown timer.
Metrics
| Model | Status | Wall time | Input | Thinking | Output | Equiv. paid cost | Finish |
|---|---|---|---|---|---|---|---|
| Gemini 4 Hydrogen | ran | 54.8s | 159 | 1,409 | 9,841 | $0.0846 * | STOP |
| Gemini 3.8 Flash | error | — | — | — | — | — | — |
| Gemma 4 31B | ran | 1m 51s | 129 | 825 | 2,976 | $0.0000 * | STOP |
* Equivalent paid cost; ran on free tier. Hydrogen $1.50 / $7.50 per 1M; 3.8 Flash $0.75 / $3.75 per 1M through 31 Dec 2026; Gemma listed free ($0.00). Output price includes thinking tokens.
What we observed
- Gemini 4 Hydrogen: A complete start screen shows score, best, a 60-second timer, a control legend, and a Start Delivery button. After start, a pixel-art cat on a bicycle rides a pier with a lighthouse and water, falling buns, and a countdown timer.
- Gemini 3.8 Flash: No output this session. The Gemini 3.8 Flash API returned HTTP 503 (“high demand”) or dropped the connection. We tried once more after the first failure and did not retry in a loop. This was not a quota / 429 error.
- Gemma 4 31B: The start screen is titled HARBOR CAT COURIER and lists WASD or arrows. Two dark rectangles overlap the “PRESS ANY KEY TO START” line; they match obstacle blocks used in play. After a key press, a 60-second game starts: a small orange cat-like figure with two wheels, a falling bun, gray blocks, a score, and a countdown timer.
Observed facts only; we did not score the models and we did not fix their code.