I Made Best AIs Rebuild 3 Real Games. One Cost 5 Times More

Two AI models got the same brief: rebuild three real games from scratch in Godot. Same screenshots, same engine, no downloaded assets — even the sound effects are generated by code. Each model could look at its own build only 10 times. After that, it was flying blind. Opus 5 (Anthropic) vs Kimi K3 (Moonshot). Steep, Hollow Knight: Silksong and Fall Guys, rebuilt from nothing. I gave each model feedback between runs and let them iterate. Some of what came back genuinely surprised me — and the difference you can't see turned out to be more interesting than the one you can. Final cost: ~$190 vs ~$37. 00:00 The brief 00:52 STEEP — Opus, run 1 01:27 Opus, run 2 — the trail problem 02:05 Opus, run 3 — this is a game now 02:28 STEEP — Kimi K3, all three runs 03:26 Steep verdict 03:34 The part you can't see: nobody used the physics engine 03:53 SILKSONG — Opus 04:43 SILKSONG — Kimi K3 05:04 Silksong verdict 05:09 FALL GUYS — both builds 05:45 Real ragdoll vs a faked one 05:52 Fall Guys verdict 06:01 What it cost 06:08 Outro Everything you see was built by the models — the mountain, the cave, the characters, the shaders and every sound effect. Nothing downloaded.