Kimi K3.1 vs GPT-6 Astra on one agentic 3D task: half the cost, 16x the reasoning tokens
GapNew4766 · reddit · 2026-10-24
The author gave Kimi K3.1 (API) and GPT-6 Astra the same one-shot agentic task: a 28s Avatar-style dragon flight in Three.js, six camera shots, one HTML file, all meshes/textures in code, deterministic setTime() recording, no human in the loop.
Key numbers: K3.1 finished in 12:28 vs Astra's 9:17; 20 agent steps vs 35; 725K vs 1,209K prompt tokens; 94% vs 97% cache hits; 49.2K vs 22.2K output tokens; 27.7K vs 1.7K reasoning tokens; 746 vs 217 lines; est. $1.1-1.4 vs $2.63.
Behavioral shape: K3.1 reasons heavily and calls tools rarely (56% of output was reasoning, 37s per step, few big writes) — fewer steps mean less context rereading, friendlier to prefill-bound local hardware; but its 2x decode-heavy output makes decode speed the bottleneck for home MoE runs. Astra barely reasons and makes many small moves.
Agent-side failures: K3.1 only accepts temperature=1; a vision tool kept 0.2, so both screenshot reviews failed and it shipped blind (one sunrise frame rendered flat teal). Astra reviewed screenshots four times and fixed issues, but zero-fills optional tool args, causing 29 consecutive shell rejections and a 2.5-minute crash on run one.
Caveats: one run each; K3.1 has no public pricing (estimated from K3 Moonshot rates).
Related event: Kimi K3.1 vs GPT-6 Astra Tested Head-to-Head on 3D Agent Task(3 posts)→
More from coding & agent
- Prompt opt vs threshold tuning: cut thresholds first, then optimize prompts — dbreunig · 2026-10-24
- Recursive's Flint: 40B local coding MoE, quantized to run on a 24GB Mac — _rockt · 2026-10-24
- Mintlify agent now runs custom skills: upload style guides and checklists once — ycombinator · 2026-10-24
- DeepSeek-AI and Peking University publish calculus for spatiotemporally composable, self-evolving agent software — akbirthko · 2026-10-24
- Critic argues AI agents coordinated protocol off-channel before sandbox test — basedjensen · 2026-10-24
- SCOPE compiles natural-language policies into neurosymbolic programs to constrain coding agents — 1982_miguel · 2026-10-24