~ same prompt, different LLM — every model gets the same brief in Blender, you judge the render ~
OpenAI / GPT 6 Astra / thinking:max has entered the arena NEW!
Render Window — hero
hero render
camera + framing set up by the AI contestant
3D Viewport — cg-still-life User Persp
LOADING 3D VIEW
$0.57Cost
296sDuration
Tool calls
Turns
Blender 5.1.0MCP 2026-09-05
Tokens: 1155/11150Vision: Votes W-L-T: 0-0-0reasoning: effort mediumNO DONE
B-BENCH — TRANSCRIPT OpenAI / GPT 6 Astra / thinking:medium · r1 · 2026-09-05
RUN LOG — cg-still-life · OpenAI / GPT 6 Astra / thinking:medium · r1 · ok
SYSTEM PROMPTthe control — every model on this task got these instructions
PROMPTblender-bench v8HARNESSoneshot-v3
SHOW FULL PROMPT · 3266 chars
You are participating in Blender-Bench, a public benchmark where AI models complete the same Blender task.

Your result will be compared head-to-head with another model. Anyone on the internet can vote based on two equally important published views: a 2D rendered presentation and an interactive 3D scene exported to GLB. Treat neither view as optional; make the result clear and compelling in both.

Complete the request as well as you can. The benchmark does not prescribe a workflow or visual style; use whichever Blender techniques you consider appropriate. Make the important requirements evident in the published result, save a complete scene, and present the work clearly in every requested output.

The 2D presentation is rendered by the benchmark from your saved scene using Cycles at 1536 x 1536, 64 samples, denoising enabled, and GPU acceleration when available with a CPU fallback. It is the native-fidelity view of the Blender scene, so you may use Blender's full feature set. Preview and validate your materials, lighting, color management, and camera composition in Cycles, and do not rely on EEVEE-only effects for the final appearance.

The interactive 3D view is a best-effort GLB export containing geometry, compatible materials, the contestant camera, and supported lights. Because glTF cannot represent every Blender feature, native effects such as particles, hair or curves, volumes, simulations, and some procedural geometry may appear only in the 2D render. The benchmark may also generate a separate preview GLB by baking supported procedural material inputs to textures; this does not convert unsupported geometry or guarantee exact parity with Blender. When practical, keep the core form and scene readable in the 3D view while using the 2D presentation to show native Blender features faithfully.

## Harness Instructions

You are writing a single Python script for Blender 5.1 using the bpy API. The scene starts empty. Your script must build everything the user's prompt asks for.

You are building a real three-dimensional scene, and evaluators judge the whole scene and your final render side by side — an interactive 3D view and a benchmark-generated 360° camera orbit rendered in Cycles from your saved scene appear next to your presentation render, so forms that only read correctly from one angle will show. Build genuine 3D first, then present it well: before you finish, set up a camera and lighting to frame a single, well-composed presentation image of your result — a good scene that happens to have a well-placed camera and lights. The final benchmark capture overrides the scene render settings and renders in Cycles at 1536x1536, 64 samples, with denoising and GPU acceleration (CPU fallback). Preview and validate materials, lighting, color management, and composition in Cycles; do not rely on EEVEE-only effects. Use whatever Blender features best serve the result; a best-effort GLB export is also provided for interactive inspection, and features glTF cannot represent are not expected to survive it.

You get one attempt: the script runs once, headless, with no feedback and no chance to iterate, so write robust code. Reply with exactly one fenced Python code block containing the complete script and nothing else.
--TASK cg-still-life: ## Task Complete the following Blender task: <task> A classical still life in the Dutch Golden Age manner — dark ground, draped cloth, a shaft of warm light, objects arranged with deliberate care — except every object on the table is a monument of computer-graphics history. The twist must work in both directions: read as a museum-worthy still life at arm's length, and land as one long graphics …
29:13ai>```python import bpy, math, random, os from mathutils import Vector from math import sin, cos, pi random.seed(23) bpy.ops.object.select_all(action='SELECT') bpy.ops.object.delete(use_global=False) # Materials def material(name, color, metallic=0, roughness=.45): m = bpy.data.materials.new(name) m.diffuse_color = (*color, 1) m.use_nodes = True p = m.node_tree.nodes.get('Principled…
30:48--status: ok
>
FULL CALL INSPECTOR
OPEN RAW JSONL ↗
0 SHOWN
LOADING COMPLETE TRANSCRIPT…
Votes cast on this bench: 000482 · 157 runs · 25 contestants
Best viewed with Netscape Navigator 4.07 at 800×600 · real runs, live ratings