Kimi K3.1 Pro vs GPT-6 Astra: Four Blender Render Tests and a Quota Limit

The scoreboard, and what it leaves out

In its October 23, 2026 launch snapshot, Artificial Analysis scored Kimi K3.1 Pro 55 on its Intelligence Index and ranked it fourth overall; Kimi K3 scored 44. This historical composite covers a different set of tasks from the Blender demonstrations below, so it cannot determine which model produces better renders.

A composite score does not tell you how many attempts a shot needs or whether a paid tier lasts a working day. The following examples summarize a third-party demonstration, not a test conducted by this site.

Test 1: a pelican on a bicycle

The prompt was the same for both models: an adult pelican riding a teal touring bike along a coastline, built in Blender as a 4-second loop, with detailed feathers, a visible throat pouch, a metal frame, and both feet tracking the pedals. The pelican had already become a 3D exercise on X, with Addy Osmani building a city ride in Three.js and Simon Willison staging one in Blender.

Seen side by side, Kimi's bird is heavier and the neck curve is more pronounced. Astra's bird is leaner, spreads its black flight feathers wider, and adds a rear rack the prompt never mentioned. Both models got the pedaling motion across, so the gap is in silhouette and surface detail rather than in whether the task was understood.

The first attempts were worse. Astra scaled an eye until it became a black sphere that covered the head. Kimi hit a modeling error, then a second error while filling in the environment.

One repair round and one art pass each produced the clips above. Feathers and legs still look procedural, though less so than in the earlier K3 output.

Test 2: a foldable phone product shot

Both models received the same Apple reference photo and the same instruction: rebuild the phone in Blender, light it, animate the hinge, and add a slight camera turn.

Kimi's first version cropped the bottom of the phone and left a visible seam in the background. Astra called the wrong Blender API, fixed that, then had to correct gray materials and highlights that were too hot.

The result works for a previz pass or a pitch deck. It does not work for ecommerce. Neither build used CAD or measured dimensions, and the lens module and hinge were simplified. A camera move exposes that quickly.

Test 3: a Roadster jump

The brief: a white sports car climbs a ramp, rolls in the air, and lands, all inside one moving shot.

Kimi's first render had broken geometry across the nose. Astra's glass reflected light like crumpled plastic. Two rounds each fixed most of the surface problems. The body stayed simplified, and a low camera angle pushed the car out of frame.

This is where the limits are easiest to see. A hero shot with a real car needs accurate proportions, convincing surfacing, and a camera move that holds across dozens of frames. Neither model got there.

Test 4: an eight-second robot short

The scene: an old projection room. A robot built from film reels sets down a canister, lifts its head, and its lens lights up. Eight seconds, one action, no cuts.

Kimi's robot is rounder, and a reel in the foreground gives the room depth. Astra added gauges and vents that were not in the brief, and its beam reads more clearly on screen.

Both robots came from scripts the models wrote. Neither imported an official model. The action is simple, but a moving clip tells you more about a character than a concept still does.

The reported quota interrupted the run

The creator reported using a 199 CNY Kimi membership and hitting its five-hour usage limit partway through the session. The remaining cases waited for a reset. Limits and plan terms may differ by account and date.

Membership prices and usage limits can change. Check the current terms for your own account before budgeting a 3D workflow.

Allow time for revisions: the examples show that a first render may need both technical fixes and an art pass.

How the renders were produced

The models wrote Python that built scenes in Blender. Blender or a browser ran the scripts and rendered the frames. Codex organized the calls and checked the output. Kimi ran through the Kimi Code interface, and Astra ran as an independent control.

This is not text-to-video. Every frame came from generated code and a render, which is why the failures look like Blender errors rather than prompt drift.

The Blender project files were not published with the test, so the renders could not be rerun independently.

Which model to pick this week

If the job is a concept frame or a scene layout, Kimi K3.1 Pro is competitive and the choice comes down to taste. If the job is a final product shot, neither model replaces a 3D artist, and the repair loop is most of the work.

Usage limits can matter as much as render quality when choosing a workflow. Check the limits on the plan you actually use.