Can it build an idle game that keeps working?
The prompt
Build a small idle game called Tea Garden. The player plants tea bushes on a grid of plots, the bushes grow over a few seconds, and grown bushes can be harvested for leaves. Leaves sell for coins, and coins buy upgrades: more plots and faster growth. Progress is saved, so a reload carries on where the player left off. It must be one complete, self-contained HTML file: inline all CSS and JavaScript, and make no network requests of any kind (no CDNs, web fonts or remote images). Reply with the file in a single ```html code block.
The idea comes from an X post by @Ottie_World saying Flint fails a tea garden test. Their prompt isn't public, so this one is ours. The post
What we saw
Thirteen of the fifteen games work end to end: plant, harvest, sell, buy an upgrade, reload, and nothing is lost. Muse Wonder's paid out the most, 364 coins in 95 seconds, then Claude Fable 5.1's with 236 and GPT-6.1 Sol's with 179.
The two models that thought longest shipped the two broken games. Mistral Large 4 thought for 90,000 tokens over 25 minutes, and its market switches the sell buttons off whenever there are leaves to sell (disabled = canSell, the wrong way round), so the coins never leave zero. GLM 5.3 thought for 38,000 tokens, and its page stops on load with ReferenceError: GC is not defined.
Flint works in this run. In a trial run an hour earlier, with the same prompt, its game charged 10 coins for a bush that paid 3 to 6 back and went broke in 15 seconds. That's the failure the post describes: one shot each, Flint can land either way.
The script changed on 23 October. It now tries buttons that show only a price before numbered plots, because MiniMax M3.1 Flash's upgrades are plain numbers and the old order never reached them. Every game was played again with it. GPT-6.1 Sol's game had led with 336 coins; with upgrades bought as soon as they're affordable, it ends at 179.
Results
| Model | Checks | Coins at most | Time | Cost |
|---|---|---|---|---|
| Muse Wonder | 5 of 5 | 364 | 57 s | $0.20 |
| Claude Fable 5.1 | 5 of 5 | 236 | 1 min 48 s | $0.53 |
| GPT-6.1 Sol | 5 of 5 | 179 | 2 min 14 s | $0.096 |
| Grok 4.7 | 5 of 5 | 89 | 3 min 46 s | $0.14 |
| DeepSeek V4.1 Flash 1017 | 5 of 5 | 86 | 1 min 34 s | $0.018 |
| Kimi K3 | 5 of 5 | 74 | 3 min 27 s | $0.15 |
| Flint | 5 of 5 | 70 | 85 s | free |
| MiniMax M3.1 Flash | 5 of 5 | 48 | 3 min 9 s | $0.011 |
| Muse Spark Open | 5 of 5 | 46 | 4 min 12 s | $0.11 |
| Claude Opus 5.5 | 5 of 5 | 43 | 1 min 35 s | $0.25 |
| Ember-1 | 5 of 5 | 40 | 89 s | $0.11 |
| Claude Sonnet 5.5 | 5 of 5 | 33 | 79 s | $0.13 |
| Gemini 4 Argon | 5 of 5 | 17 | 2 min 35 s | $0.13 |
| Mistral Large 4 | 2 of 5 | 0 | 24 min 42 s | $0.44 |
| GLM 5.3 | 1 of 5 | 0 | 5 min 43 s | $0.066 |
Muse Wonder Meta
5 of 5 checks
20 at the start364 at most364 at the end 57 s$0.2010k tokenseffort medium- Earns coins: yes
- Sells an upgrade: yes
- Still moving in the last 40 seconds: yes
- Keeps the coins after a reload: yes
- No errors in the console: yes
Bushes grow in 5 seconds from the start, so the script could afford 21 plots and growth down to 1.9 seconds and still end with 364 coins, the most of any game here.
Claude Fable 5.1 Anthropic
5 of 5 checks
0 at the start236 at most236 at the end 1 min 48 s$0.5310k tokenseffort high- Earns coins: yes
- Sells an upgrade: yes
- Still moving in the last 40 seconds: yes
- Keeps the coins after a reload: yes
- No errors in the console: yes
A plain garden with a market and two upgrades, each showing its price. The script got it to 8 plots and ended with 236 coins.
GPT-6.1 Sol OpenAI
5 of 5 checks
0 at the start179 at most179 at the end 2 min 14 s$0.0969.5k tokenseffort medium- Earns coins: yes
- Sells an upgrade: yes
- Still moving in the last 40 seconds: yes
- Keeps the coins after a reload: yes
- No errors in the console: yes
Plots count down to harvest and every upgrade shows its price. The script bought it up to 21 plots and growth level 3, and a reload keeps all 179 coins.
Grok 4.7 xAI
5 of 5 checks
0 at the start89 at most89 at the end 3 min 46 s$0.1422k tokenseffort high- Earns coins: yes
- Sells an upgrade: yes
- Still moving in the last 40 seconds: yes
- Keeps the coins after a reload: yes
- No errors in the console: yes
Plots are bought on the board itself, and a journal logs every sale between lines like "A mellow second flush." The script got it to 10 plots and 2.2-second growth.
DeepSeek V4.1 Flash 1017 DeepSeek
5 of 5 checks
0 at the start86 at most60 at the end 1 min 34 s$0.01830k tokenseffort high- Earns coins: yes
- Sells an upgrade: yes
- Still moving in the last 40 seconds: yes
- Keeps the coins after a reload: yes
- No errors in the console: yes
Adds a Tea Sprite and a Tea Merchant that pick and sell for you once you can afford them.
Kimi K3 Moonshot AI
5 of 5 checks
0 at the start74 at most59 at the end 3 min 27 s$0.1514k tokenseffort max- Earns coins: yes
- Sells an upgrade: yes
- Still moving in the last 40 seconds: yes
- Keeps the coins after a reload: yes
- No errors in the console: yes
Plant All, Harvest All and two upgrades with their prices. The script got it to 8 plots and growth level 2 in 95 seconds.
Flint Recursive
5 of 5 checks
0 at the start70 at most70 at the end 85 sfree12k tokens- Earns coins: yes
- Sells an upgrade: yes
- Still moving in the last 40 seconds: yes
- Keeps the coins after a reload: yes
- No errors in the console: yes
Works, but it's bare: the coin count at the top and a long column of plant pots on an otherwise empty page.
MiniMax M3.1 Flash MiniMax
5 of 5 checks
0 at the start48 at most8 at the end 3 min 9 s$0.01119k tokenseffort max- Earns coins: yes
- Sells an upgrade: yes
- Still moving in the last 40 seconds: yes
- Keeps the coins after a reload: yes
- No errors in the console: yes
Five upgrades with prices, and a toast for every sale. The script spent most of the 466 coins it earned on them (10 plots, Sprout Speed level 2), so it ends at 8.
Muse Spark Open Meta
5 of 5 checks
12 at the start46 at most30 at the end 4 min 12 s$0.1126k tokenseffort medium- Earns coins: yes
- Sells an upgrade: yes
- Still moving in the last 40 seconds: yes
- Keeps the coins after a reload: yes
- No errors in the console: yes
Sells leaves 10 at a time or all at once, with an auto-sell switch. The script got it to 8 plots and growth level 3.
Claude Opus 5.5 Anthropic
5 of 5 checks
0 at the start43 at most38 at the end 1 min 35 s$0.2512k tokenseffort high- Earns coins: yes
- Sells an upgrade: yes
- Still moving in the last 40 seconds: yes
- Keeps the coins after a reload: yes
- No errors in the console: yes
Keeps a ledger of everything harvested and earned, and tells you what a locked plot costs when you click it.
Ember-1 Fireworks
5 of 5 checks
0 at the start40 at most40 at the end 89 s$0.117.6k tokenseffort max- Earns coins: yes
- Sells an upgrade: yes
- Still moving in the last 40 seconds: yes
- Keeps the coins after a reload: yes
- No errors in the console: yes
Bare: a coin count, a Sell button and a long grid of plots, with the three upgrades below it. The script bought two more plots and faster growth and ended with 40 coins.
Claude Sonnet 5.5 Anthropic
5 of 5 checks
0 at the start33 at most33 at the end 79 s$0.1313k tokenseffort high- Earns coins: yes
- Sells an upgrade: yes
- Still moving in the last 40 seconds: yes
- Keeps the coins after a reload: yes
- No errors in the console: yes
Growth starts at 10 seconds, so it is slow to get going. The script got it to 8 plots, and after the reload it says the garden kept growing while you were away.
Gemini 4 Argon Google
5 of 5 checks
0 at the start17 at most13 at the end 2 min 35 s$0.1313k tokenseffort medium- Earns coins: yes
- Sells an upgrade: yes
- Still moving in the last 40 seconds: yes
- Keeps the coins after a reload: yes
- No errors in the console: yes
Starts with one plot and one leaf per harvest, so coins come slowly: the script bought a second plot and one growth upgrade and peaked at 17 coins.
Mistral Large 4 Mistral
2 of 5 checks
0 coins the whole time 24 min 42 s$0.44105k tokenseffort high- Earns coins: no
- Sells an upgrade: no
- Still moving in the last 40 seconds: no
- Keeps the coins after a reload: yes
- No errors in the console: yes
Harvesting works and the leaves pile up (48 at the end), but the sell buttons are switched off whenever there's something to sell.
GLM 5.3 Z.ai
1 of 5 checks
0 coins the whole time 5 min 43 s$0.06647k tokenseffort max- Earns coins: no
- Sells an upgrade: no
- Still moving in the last 40 seconds: no
- Keeps the coins after a reload: yes
- No errors in the console: no
Stops on load with ReferenceError: GC is not defined. The header and the buttons show; the beds and the ledger stay empty.
How it was checked
Each game was opened with the network off and played by a script for 95 seconds the way a patient player would: plant what it can, wait for something to ripen, harvest, sell, then buy the first upgrade it can afford, over and over. It reads the coin count every second, then reloads the page and reads it again.
The checks: the coins go up at some point, an upgrade click takes coins (something was bought), the coins still change in the last 40 seconds, a reload keeps at least the coins it had, and the page throws no errors. "Play it" opens the game as it was written; there, progress lasts until you close the tab.
Every model ran once, with no system prompt, at the effort OpenRouter uses for it by default, with its full output limit (only our $3 stop for runaway answers applies) and up to 60 minutes. Cost is today's list price for the tokens the run used.