M3.1-Flash is the worst performer, and gpt-6-astra is astonishingly bad at this, only M3.1-Flash was worse
Let me know in DMs if you want me to test more models on this private gamefeel/gamedev benchmark.
Only tested one subsection of Opus 5.5 on this as I ran out of credits.
M3 traffic got wild, so we shipped overnight.
Inference serving upgraded at 22:00 Beijing / 7:00 AM PT.
TPS much smoother now. Most users should be seeing 50–70 TPS.
gave M3.1 Flash one prompt
build the Titanic wreck as an explorable 3D scene in one html file
thought it would just make a basic ship scene but somehow it added a pilotable sub , sonar , marine snow , silt clouds , fish and the wreck split into bow and stern
even added a ghost mode where you can see the original ship over the wreck
best part is flying around with the spotlight and kicking up silt near the seabed
this is also where M3.1 Flash makes sense
it’s fast , handles pretty complex coding + interaction logic well , and the cost is low enough that i can keep changing things and rerunning it without thinking too much about every iteration
the model is also open-weight now
pretty fun test