TL;DR

Across four Blender shots run from matching prompts, Kimi K3.1 Pro and GPT-6 Astra produced work in the same bracket. Modeling and motion sat close; outlines and surface finish did not. All four failed their first attempt. Kimi's benchmark: 55 on the Intelligence Index from Artificial Analysis, version 4.3.2. It ranks fourth overall and leads the Chinese field, with GPT-6 Astra behind it in that snapshot. K3 managed 44. A 5-hour usage cap on a 199 CNY membership stopped the session before it finished.

All four first attempts failed

Repair work showed up on every brief. The specific breakages are the useful part, because they mark where this workflow gives way first.

Astra let the pelican's eye grow until it read as a black ball covering the head. Kimi stopped on a modeling fault, then stopped again while building the surroundings.

The phone brief split them. Kimi's opening build chopped the lower edge of the device and left a background seam. Astra picked the wrong Blender API call, repaired it, then still had flat gray surfaces and burned highlights.

Kimi's Roadster nose arrived with broken geometry. Astra's glazing caught light like creased plastic.

Cost of repair differed: a modeling patch and an art pass on the bird, two rounds on the car, a separate pass on the phone to get the API call working.

BriefKimi K3.1 Pro outputGPT-6 Astra output
Bird on a teal tourer, 4-second cycleDenser body, a deeper bend at the neck; one modeling fault, one art passSlimmer build, black primaries spread wider, a rear rack nobody asked for; eye scaling corrected
Foldable handset, hinge moveHandset base cut off, plus a background seam, on the opening attemptWrong Blender call, flat materials, blown highlights on the opening attempt
White Roadster ramp leap, 6-second cutBroken geometry at the nose, resolved across two roundsGlazing looked like creased plastic, resolved across two rounds
Robot in a projection room, 8 secondsSofter figure, foreground reel carrying the depthAdded dials and vents, brighter beam

Where they diverge

Both models got the same description: a grown pelican pedaling a teal touring bike beside a coastline, looping four seconds, feathers drawn in detail, throat pouch on show, metal frame, both feet locked to the pedals.

Line up the two bird clips and mass is the first difference. Kimi's pelican carries more weight and folds its neck deeper. Astra's is slimmer, spreads its black primaries wider, and bolts on a rear rack that nobody ordered. Pedaling reads correctly in both. What separates them is outline and surface finish, not comprehension of the brief.

Plumage and leg meshes still betray their procedural origin on both sides, a little less than K3 managed earlier.

Kimi K3.1 Pro, four seconds looping after a modeling patch and one art pass. More mass through the body, a deeper neck bend.
Kimi K3.1 Pro Blender animation of a pelican pedaling a teal touring bike beside a coastline in a four-second loop
GPT-6 Astra: identical prompt and repair routine, slimmer outline, plus a rear rack nobody requested.
GPT-6 Astra Blender animation of the same pelican and bicycle brief, slimmer body and an extra rear rack

The setting was a projection room past its prime, occupied by a figure welded out of film reels. It puts a canister down, raises its head, and switches its lens on. Eight seconds, single action, no editorial cuts. The two models read it differently. Kimi builds a rounder character and drops a reel into the foreground to open up depth. Astra invents dials and vents the brief never asked for, and its beam also reads brighter on screen. Neither brought in a ready-made rig; both grew from code the models authored. Kimi's clip shows six of the eight seconds.

Kimi K3.1 Pro: six of eight seconds, with the front reel putting depth into the room.
Kimi K3.1 Pro Blender animation of a robot made from film reels lifting its head as its lens switches on
GPT-6 Astra: same room, extra mechanical parts, a stronger beam.
GPT-6 Astra Blender animation of the same film-reel robot scene with added mechanical parts

For the handset, both started from one Apple press photo and the same four jobs: rebuild it in Blender, light the scene, run a hinge animation, and nudge the camera through a small turn.

Hinge sequence for Kimi K3.1 Pro, rebuilt from the press photo.
Kimi K3.1 Pro Blender animation of a foldable phone opening at the hinge, modeled from an Apple press photo
GPT-6 Astra: identical photo and shot list.
GPT-6 Astra Blender animation of the same foldable phone opening at its hinge

What the clips cannot support

Ecommerce is out first. The handset animation survives a previz review or a pitch deck and nothing more. No CAD model or measured dimension went in, the lens block and hinge stayed approximate, and any camera move gives that away.

The Roadster breaks on other grounds. Its brief asked for a white sports car climbing a ramp, rolling in the air, and landing inside one continuous shot. Making that look like a real car demands proportions that survive a close look, reflections that convince, and a move that stays stable for dozens of frames. Neither cleared that bar. Two rounds left the body simplified, and a low angle also cropped the car.

Kimi K3.1 Pro, six seconds following two rounds of revision.
Kimi K3.1 Pro Blender animation of a white sports car launching off a ramp, rotating midair and landing
GPT-6 Astra: same six seconds, same two rounds.
GPT-6 Astra Blender animation of the same white Roadster launch, midair rotation and landing

The score, and what it cannot answer

Kimi K3.1 Pro shipped in October 2026, and most readers opened with the leaderboard. It holds up: Artificial Analysis's Intelligence Index, version 4.3.2, gives it 55, built from 10 evaluations spanning terminal work, software engineering, and reasoning. That puts it fourth overall and top of the Chinese field, past GPT-6 Astra in that snapshot. K3 sat at 44.

Composites do not report how many passes one shot takes, nor whether a purchased tier survives eight hours. Those were the two questions driving this test.

The session hit a 5-hour ceiling

The account behind the test was a 199 CNY membership. Its five-hour allowance ran out before the last cases, which then waited for the reset.

That ceiling matters on 3D work. Monthly pricing for the product now runs 99 to 699 CNY, and iteration at this density drains a tier. WorkBuddy also sells K3.1 Pro against credits, cheaper than K3.

Plan on three or four rounds for each shot. Nobody gets a finished 3D clip from a single prompt today.

Trying several models inside one week gets costly, and some plans do not repay their price. The multi-model coding plan we stock puts eight text models behind one key, with pooled points, billing by month or once, and agent setup docs for seven tools. Stock reflects what we have.

How each frame reached the screen

Code written by the models assembled the scenes in Blender. Either Blender itself or a browser ran that code and produced frames. Codex scheduled the calls and reviewed the output. Kimi used Kimi Code as its interface; Astra acted as a separate control.

Nothing here is text-to-video. Generated code plus a render produced every frame, so a miss surfaces as a Blender fault rather than a prompt drifting.

No project files accompanied the test, which means these clips cannot be reproduced independently.

Choosing between them

A concept still or a scene layout leaves Kimi K3.1 Pro competitive, the rest is taste. A finished product shot still needs a 3D artist; neither model removes that person, and revision takes most of the hours.

Treat the cap as part of the budget. Losing the afternoon at hour four is the real risk.