Blender-Bench — which robot blends best?

blender-bench.org — blind vote · space-cowboy-duel

Blender-Bench V1.0names hidden until you vote

VIEW BOTH TO VOTE — open Variant B

keys: ←/1 A wins · 2/→ B wins · T/3 tie · X/4 both bad · S skip

Render Window — hero

camera + framing set up by the AI contestant

Render Window — hero

camera + framing set up by the AI contestant

TASK

◀all▶

RATE THE DETAILS — optional

beyond the overall pick — call individual wins.

Lighting

which is lit better?

Composition

framing & composition

Modeling

geometry & shapes

Materials

materials & surfaces

Render

final image quality

Adherence

matches the prompt?

lens rounds — occasionally answer for one dimension

TOP OF THE BENCH

1 OpenAI / GPT 6 Astra / thinking:max2224

2 OpenAI / GPT 6 Astra / thinking:medium1986

3 OpenAI / GPT 5.6 Sol / thinking:max1777

FULL LEADERBOARD ▶

26CONTESTANTS24TASKS163RUNS1671VOTES

*** NEWS ***

13.jul.26gpt-5.6-luna enters the arena NEW!
13.jul.26cabin-pond complete — first head-to-head pairs are live
10.jul.26one-shot mode approved: headless renders coming
05.jul.26blender-bench is born. two tasks, one cube, big dreams

HOW IT WORKS

1LLMs drive a real Blender through MCP tool calls — no human hands.

2Every run gets standardized renders + the model's own hero shot.

3You vote on blind pairs; Bradley–Terry turns votes into the ranking.

*** COMING SOON ***

Second Wave SOON

TASK SUITE

The current tasks ask: given a blank scene, can a model make something beautiful? The Pipeline suite asks the question every production artist cares about — given someone else's file and a concrete job, can it do real Blender work?

Retopology — clean quad mesh over a dense sculpt
UV unwrapping + PBR texturing on an untextured prop
Rigging — armature, weights, working deformation
Rig animation — a looping walk cycle on a provided rig
Scene cleanup — naming, hierarchy, orphan data

3D Bench SOON

SIBLING BENCH

A sibling benchmark for the other ways models make 3D — no Blender required. Same blind-vote format, pointed at code that draws.

three.js scene generation
p5.js sketches + generative scenes
Judged by the same pairwise human voting

BYOK SOON

RUN YOUR OWN

Bring your own key. Point your own API key at a task and run a model through the bench yourself, then drop the result into the arena.

Your key, your model, your run
Any task in the library
Results feed the same leaderboard

★ Source on GitHub SOON · ♥ Support the bench

CURRENT STANDINGS

#	CONTESTANT	RATING	RUNS	DONE%
1	OpenAI / GPT 6 Astra / thinking:max	2224	1	0
2	OpenAI / GPT 6 Astra / thinking:medium	1986	16	13
3	OpenAI / GPT 5.6 Sol / thinking:max	1777	4	75
4	Anthropic / Opus 5 / thinking:medium	1766	5	80
5	OpenAI / GPT 5.6 Sol / thinking:medium	1717	2	50
6	Moonshot AI / Kimi K3 / thinking:max	1706	3	33
7	OpenAI / GPT 5.6 Sol	1683	2	100
8	Anthropic / Claude Fable 5.1 / thinking:medium	1676	1	100
9	OpenAI / GPT 5.6 Luna / thinking:max	1665	20	95
10	OpenAI / GPT 5.6 Terra	1558	2	100
11	OpenAI / GPT 5.6 Luna	1548	35	46
12	Anthropic / Claude Fable 5 / thinking:medium	1508	1	100
13	Google / Gemini 3.6 Flash / thinking:medium	1508	2	0
14	Google / Gemini 3.8 Flash / thinking:medium	1389	3	0
15	DeepSeek / Deepseek V4 Flash 0731 / thinking:max	1386	6	17
16	Google / Gemini 3.5 Flash / thinking:medium	1357	11	18
17	DeepSeek / Deepseek V4 Pro	1335	3	100
18	Z.AI / GLM 5.2	1326	5	60
19	Anthropic / Sonnet 5	1324	3	0
20	Ox Alpha / thinking:max	1315	1	0
21	Anthropic / Sonnet 5 / thinking:low	1296	8	38
22	DeepSeek / Deepseek V4 Flash	1179	11	36
23	Google / Gemma 4 31b It	1148	10	70
24	Anthropic / Haiku 4.5 / thinking:xhigh	1133	3	67
25	Google / Gemini 3.1 Flash Lite	991	3	100
26	Ox Alpha / thinking:low	—	2	0

★ Source on GitHub SOON · ♥ Support the bench

Votes cast on this bench: 001671 · 163 runs · 26 contestants

Best viewed with Netscape Navigator 4.07 at 800×600 · real runs, live ratings