Opus 5 Tested: Major Leap in 3D Game Generation, but Costs Nearly $1K
sebnadeau · reddit · 2026-07-30
A developer tested Claude Opus 5 using the WorldBuild Bench, a harness that evaluates LLMs by having them build playable 3D games to test spatial, temporal, and causal coherence.
- Quality: Opus 5 shows a clear step up in 3D modeling, texturing, effects, and lighting, outperforming the previous standout Fable 5. It suggests models are entering a phase where they can create all necessary game content from scratch.
- High Costs: Opus 5 is the most expensive model tested yet. Across three game briefs (racing, arena combat, physics puzzle), it cost a total of $931.88 (avg. $310 per run), surpassing Fable 5's already high costs.
- Generation Time: Runs average about 8 hours. Opus 5 takes longer to call a task 'done', continuously iterating and spawning 13-15 subagents per run to refine its output.
The author concludes that while Opus 5 reaches conclusions slower, its extended iteration and better understanding of outputs directly translate to higher quality results.
Related event: Opus 5 Impresses in 3D Game Generation at High Cost(2 posts)→
More from coding & agent
- 'LLM code is sloppy' rebutted with real ffmpeg code: identical if/else branches in a top library — kingslyj · 2026-10-16
- Office-editing tool calls need file, worktree and unit: lessons from dsh-univer-office — FunDistribution6537 · 2026-10-16
- Building an acceptance harness for AI spreadsheet edits: allowlists, recalc checks, idempotency — ukdicha-modakkk · 2026-10-16
- Agent actions need checks outside the model: NVIDIA demos lean on guardrails amid wave of real agent incidents — Creamy-And-Crowded · 2026-10-16
- Vitalik builds "paranoid mode" agent wrapper with Tor, zkAPI and local prompt rewriting — DavideCrapis · 2026-10-16
- Agent fine-tunes Qwen3.5-0.8B with synthetic data: 4.7%→76.3% accuracy, 11.9× faster — ben_burtenshaw · 2026-10-16