Kimi K3.1 vs GPT-6 Astra head-to-head on an identical agentic 3D task
GapNew4766 · reddit · 2026-10-24
The Atomic Agent developer ran Kimi K3.1 Pro and GPT-6 Astra on the exact same prompt and tools to independently build a 28s Avatar-style Three.js cinematic, one run each, no retries.
Key numbers:
- Wall time: Kimi 12:28 vs Astra 9:17; steps: 20 vs 35
- Context read: Kimi 725K (94% cache hits) vs Astra 1.2M (97%)
- Output: Kimi 49.2K tokens (27.7K reasoning, 56%) vs Astra 22.2K (only 1.7K reasoning)
- Code: Kimi 746 lines/50KB vs Astra 217 lines/36KB
- Cost: Kimi $1.1-1.4 (estimated; no K3.1 pricing yet) vs Astra $2.63
Style contrast: Kimi thinks long, takes few steps, and writes more code — but built blind: a client bug broke its vision calls, so it shipped without seeing screenshots, missing a flat teal sunrise frame. Astra took 6 screenshots, acted on visual feedback, and produced tighter, more consistent shots.
The author admits two bugs made it not fully fair for Kimi (Astra's first run crashed in 2.5 minutes and was rerun; Kimi's vision call kept a bad temperature). Disclosure: author works on Atomic Agent.
Related event: Kimi K3.1 vs GPT-6 Astra Tested Head-to-Head on 3D Agent Task(3 posts)→
More from coding & agent
- Enable Gemini's agentic video understanding for ~90% lower token usage and better accuracy — MickeySteamboat · 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
- Learn Go with Tests: 399-page open TDD textbook now readable with an AI tutor — burkov · 2026-10-24