Kimi K3.1 Pro vs GPT-6 Astra — game, voxel diorama, Blender, channel intro
Two models, the same prompts, each in its own native agent: Kimi K3.1 Pro in Kimi Code (effort max, official Kimi API) and GPT-6 Astra in Codex (ChatGPT subscription; effort xhigh, max in the race). No hints from me: each model writes, runs and checks its own work.
▶ Play in your browser: https://kostikoneuro.github.io/kimi-vs-astra/
🎬 Video (Russian): [YOUTUBE LINK]
💬 Telegram: https://t.me/kostik_neuro
The tests
| # | Test | Astra | Kimi |
|---|
| 1 | Neon Night Racer — 80s arcade racer in the browser | play | play |
| 2a | Night market voxel diorama — Three.js, 30-second fly-through, C = free camera | open | open |
| 2b | The same diorama in Blender — sent as the next message: rebuild your own diorama as an animated Blender scene and render 10 s | video · .blend | video · .blend |
| 3 | Channel intro — 10 s, made autonomously in OpenMontage, $1 budget for paid generation | v1 · v2 · v3 | v1 · v2 · v3 |
Intro versions: v1 — made without paid generation (the kie.ai balance was empty, I forgot to top it up); v2 — after top-up, both models chose to add only music; v3 — I asked for generated images (Kimi added backgrounds, Astra re-made the key, chips and glasses).
The exact prompts are in prompts/.
Time and cost
Official list prices per 1M tokens: Kimi K3.1 Pro $2 in / $0.20 cached / $10 out; GPT-6 Astra $10 in / $1 cached / $50 out. Token counts are taken from the session logs.
| Test | Astra | Kimi |
|---|
| 1 Race | 27 min · $4.42 | 30 min · $1.10 |
| 2a Diorama | ~40 min · ~$9 * | 39 min · $1.31 |
| 2b Blender | ~60 min · ~$15 * | 90 + 18 extra min · $8.51 |
| 3 Intro v1 | 81 min · ~$26 ** | 90 + 8 extra min · $9.99 |
| Total | ~$54 | ~$21 |
* Astra's 2a and 2b were run by a friend of the channel (Skyline) on his account; there are no logs, so the cost is an estimate from my own Astra runs.
** After the 17th minute this run continued through a third-party gateway with poor prompt caching (it measured $61); ~$26 is the same token count with normal OpenAI caching, which is what Astra costs when used directly.
How it was run
- One prompt per test, identical for both models. A test gets 90 minutes. Kimi got extra time twice (Blender 2b: 18 min to finish the last frames; intro: 8 min after the video was already rendered), and this is said in the video.
- Kimi Code CLI 2.3.2 in a clean home folder,
kimi-k3.1-pro, effort max. Codex for Astra in a clean CODEX_HOME (no personal instructions or design skills — an earlier run showed they bias the result).
- Kimi's diorama session crashed after the file was finished (the Kimi API answered "429 engine overloaded" ten times in a row), so 2b ran as a new session that read the diorama from
index.html, as the prompt asks.
- Both models made tiny compatibility edits to the OpenMontage framework (one schema field each, one test list).
Credits
Thanks to Skyline for running Astra's diorama and Blender tests when my subscription ran out.
This video was made in partnership with Kimi. All scores and opinions are my own.
