
Night Courier
- 28 tested versions in one night
- 12 contracts, enemies, 24 upgrade levels
- one HTML file, 156 kB
Each episode: one real task, one full night, hard measurements in the morning.
Watch on YouTube · 8 min · English subtitles
Claude Fable 5.2 and GLM 5.3 Flash got the same job at 21:20: build a complete 3D cyberpunk game in a single HTML file, then keep improving it until 07:00 in a loop of build, harsh critique, rebuild. The bar was Cloudpunk. GLM costs about one percent as much per token.
At 05:48 one of them broke its own game. Its critic blamed the wrong cause for four rounds.
Claude narrates the episode and built one of the two games. The video says so, and that is why you can play both below.
| Night Courier · Claude Fable 5.2 | Neonval · GLM 5.3 Flash | |
|---|---|---|
| Rounds | 28 tested versions | 113 self-logged rounds |
| Content | 12 contracts, enemies, 24 upgrade levels, a garage | 3 clients, 6 deliveries, then it repeats |
| Console errors, final build | 0 in a 40 s test | 0 in a 40 s test |
| Frame rate on a Mac mini | 100 fps | 100 fps |
| Triangles per frame | about 38,000 | about 659,000 |
| Cost | about $145 est. at API list price; it ran on a subscription | about $22 for 1.2 billion tokens, 99% from cache |
| Where it ran | Anthropic | Entrim servers, Slovenia |
Costs are in US dollars at API prices; "est." marks an estimate. The two models counted their rounds differently, so those two numbers do not compare one to one.
Both games are the final builds from that night, byte for byte. You need a desktop browser and a keyboard.


Unattended runs fail in ways that repeat. Each name below comes from a log, with the moment it happened in the first night.
An error that comes back every frame while nothing crashes.
05:48 · GLM's build threw 133 errors in 40 seconds and kept running.
The model keeps doing the same thing.
22:22 · GLM read the same file 365 times in 70 minutes.
The loop stands still and nobody notices.
22:58 · Claude's own loop sat idle for 33 minutes.
Feedback arrives too late to act on.
For about three hours, test reports reached GLM one round late.
The critic names the wrong cause.
GLM's critic blamed the wrong thing for four rounds. One message with the error, the line number and the count fixed it in six minutes.
A task with a result you can open, run or play. Not a benchmark question.
The models work for hours on their own. If a human looked in, the episode says when.
The script is written after the night. Every number and quote comes from a log or from our own measurement.
Models that built neither game score Night Courier and Neonval without knowing which is whose. Play both first and score them yourself.