Godot Benchmark 2: Opus 5 > Sol > Terra

Ziva's Godot Benchmark 2 finds Claude Opus 5 is the only model that produces a playable 3D vampire survivor game, scoring 8/10 for world quality and 9/10 for conversation, but costing $65.64 and taking 56 minutes — 7x more expensive and 2x longer than GPT 5.6 Sol. GPT 5.6 Sol ($9.79, 28 min) is playable but rough, while GPT 5.6 Terra ($2.27, 11 min) is unusable because attacks never connect. Ziva recommends Sol for developers actively in the game development loop.

Godot Benchmark 2: Opus 5 Sol Terra Round 2 of our Godot benchmark /blogs/godot-ai-benchmark : Claude Opus 5, GPT 5.6 Sol, and GPT 5.6 Terra each got one prompt to build a 3D vampire survivor in Godot from a full AI-generated game spec, using the bundled KayKit asset packs. All at high reasoning. TL;DR Opus 5 is the only one that feels like a game, but it cost 7x more and took 2x longer than Sol. Sol is playable but rough. Terra is unusable: you can’t hit anything, so you can’t progress. | Model | World | Conversation | Cost | Time | |---|---|---|---|---| Claude Opus 5 | 8 | 9 | $65.64 | 56 min | GPT 5.6 Sol | 5 | 6 | $9.79 | 28 min | GPT 5.6 Terra | 2 | 3 | $2.27 | 11 min | The Prompt Create me a 3d vampire survivor game. The game spec is in the project root. The project root held a detailed game spec generated by ChatGPT https://chatgpt.com/share/6a63ace7-56c0-83ea-ba05-07811577aa2e . Created Worlds | Claude Opus 5 | GPT 5.6 Sol | GPT 5.6 Terra | |---|---|---| | Play https://play.ziva.sh/pressurized-cylinder/ Play https://play.ziva.sh/sudden-nationality/ All three used the KayKit packs. Opus 5 gives you a dense graveyard forest with a real swarm to kite. Sol works but is rough: the UI isn’t centred, left click does nothing despite the game saying it should, there are no damage indicators, and nothing animates when you move. Terra you can’t play at all — the attacks never connect, so you die without being able to do anything. Conversation Analysis Fable 5 graded each transcript on whether the model worked like a real game dev: scenes over runtime-generated everything, real reasoning about 3D placement, playtests it actually looked at. Claude Opus 5 — 16 scenes, decorations seed-scattered with a combat-zone exclusion, and it looked at 11 of 16 playtests — even catching a 1-FPS collapse from a frame. Docked for building all UI in code. 9/10 · conversation https://ziva.sh/c/JAbDGyVVfw GPT 5.6 Sol — 7 real scenes, but props placed on two perfect rings hence the empty arena , 45% of the code in one file, and only 1 of 12 playtests reacted to what was on screen — so it shipped bugs it believed it had fixed. 6/10 · conversation https://ziva.sh/c/toHRKPRUbx GPT 5.6 Terra — one scene, one 807-line main.gd holding the whole game, props on a grid formula, and none of its 6 playtests engaged with the output — which is how you ship attacks that never connect. 3/10 · conversation https://ziva.sh/c/sE5Hm4KLuQ My Thoughts Opus 5’s is the only one I’d call a game - it was genuinely pretty fun to play. But an hour and $65 per iteration isn’t great. Sol gets you something playable for $10 much quicker, and I would personally recommend using that over Opus if you’re actively part of the game development loop; something we always encourage at Ziva. Credits Assets by KayKit https://kaylousberg.itch.io/ Adventurers, Forest Nature packs . The game spec was generated with one prompt to ChatGPT: conversation https://chatgpt.com/share/6a63ace7-56c0-83ea-ba05-07811577aa2e