Kimi K3.1 Pro 3D Benchmark: Temple of Heaven in Blender, Robot Motion, and the Cost of Each Run

First, the showpiece: a temple assembled by scripts

Verdict: nothing else in the round matched it for completeness. It was also the slowest.

Given one reference photo, K3.1 Pro opened Blender, authored Python inside it, ran that code, and returned an exploded view of the building's structure rather than a handsome hollow shell.

Everything that makes the Temple of Heaven readable survived the trip: three tiers of eaves, the round footprint, dark blue glazed tiling, vermilion doors and windows, decorated beams, the white marble terrace, the golden finial. The orbit and interaction design wrapped around the model was the last thing to expect from something that only writes scripts.

Two footnotes. Generation drags. And the operator's side build, a Shenzhen metro station modeled halfway in Blender with the lighting untouched, suggests the geometry scales better than the wait does.

Motion, where most models fall over

Verdict: every motion landed.

Unitree's G1 set the reference here. Humanoid figures punish small mistakes: dozens of joints, rigid surfaces, and a kinematic chain where one wrong offset reads instantly. The K3.1 Pro output pulled off a boxing combination, a backflip, and a wave, with joints that move like they were planned rather than interpolated.

DeepSeek V4.1 supplied the comparison. It is a Flash-class model sold on light duties, yet it produced a watchable humanoid animation in a single attempt, which says more than its positioning does.

The one direct fight: two parks, one clipped castle

Verdict: close enough to call a draw.

K3.1 Pro generated its park natively in Three.js. Castle, rides, water, illumination, surfaces: all procedural. The camera behaves like a visitor's, with free orbit, saved viewpoints, and zoom and rotate, and the grounds read as one connected place instead of props on a grid.

Astra sculpts the prettier park at first glance, with spires that stack more detail. The flaw hides until you orbit around: coaster rails run right through the castle walls. On the K3.1 Pro side the surfaces are plainer and the buildings simpler, yet nothing intersects, the layout makes sense at a glance, and its water clearly beats Astra's.

What the meter recorded

The harness totals every session, and two match-ups are worth quoting. K3.1 Pro closed at $2.87 where Astra reached $6.47 on comparable work. Against DeepSeek V4.1, K3.1 Pro landed at $5.72 while DeepSeek settled at $1.87.

Read those as direction. No run published token counts, retry counts, or scoring rules, and one extra-long reasoning chain can swing a total as hard as switching models.

Where the friction sits

• Speed. Waiting on a render becomes the limiting factor before quality does.
• Access limits. The tester hit 429 responses and a subscription quota ceiling, and reports that Kimi has owned the problem.
• Surfaces. Astra keeps its edge on fine detail and small components.
• Runaway cost. Subagent sprawl and marathon reasoning chains bill faster than the scenes themselves justify.

Which model belongs where

Reach for GPT-6 Astra when the job is a large scene that must render stably and carry fine parts; its park proved that even with track through the castle. Reach for K3.1 Pro when structure, procedural builds, and interaction logic lead the requirements; its Blender temple was the most finished artifact of the round. DeepSeek V4.1 is the budget option, and the tester still flags the widest visual front-end shortfall.

One difference benchmarks tend to skip: K3.1 Pro operates Blender, SolidWorks, and Rhino through their native scripts, so no MCP connector sits in the middle. Anyone shipping real models rather than screenshots loses a layer of glue code, and that shifts a workflow further than any marginal quality gap would.

The tester's own summary puts K3.1 Pro at the open-source frontier for front-end work. Across these scenes that reading fits structure and interaction. Materials tell a different story.

Scoreboard