Kimi K3.1 vs GPT-6 Astra on the same agentic 3D task: full token and cost breakdown

GapNew4766 · reddit · 2026-10-24

An open-source agent developer ran the same one-shot task on Kimi K3.1 and GPT-6 Astra: a 28s Avatar-style dragon flight in Three.js with six camera shots, all meshes in code, deterministic frame-by-frame recording.

Key numbers

Structural insight: K3.1 reasons heavily and writes in big chunks — fewer steps means less context re-reading, friendlier to prefill-bound local hardware; but its 2x decode-heavy output means decode speed decides whether local MoE runs take 15 minutes or an hour.

Gotchas: K3.1 only accepts temperature=1; the agent's image-review sub-calls kept 0.2, so both screenshot-review attempts failed and it shipped blind (a sunrise rendered as a flat teal frame). Astra filled every optional tool argument with zeros, causing 29 consecutive shell-tool rejections and a 2.5-minute first-run crash.

Practical tip: to run K3.1 as a local agent, pin temperature to 1 across your whole stack (vision, summaries, sub-calls), not just the main chat call.

Related event: Kimi K3.1 vs GPT-6 Astra tested on same 3D agent task(3 posts)→

Original post →

More from coding & agent

coding & agent channel →