~ 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 — corgi-beach-ball User Persp
LOADING 3D VIEW
$0.03Cost
377sDuration
6Tool calls
7Turns
Blender 5.1.0MCP 1.22026-07-15
Tokens: 66667/8589Vision: Votes W-L-T: 1-6-0DONE ✓
B-BENCH — TRANSCRIPT Google / Gemma 4 31b It · r1 · 2026-07-15
RUN LOG — corgi-beach-ball · Google / Gemma 4 31b It · r1 · ok
SYSTEM PROMPTthe control — every model on this task got these instructions
PROMPTblender-bench v7HARNESSraw-v2
SHOW FULL PROMPT · 2016 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 only on the published visual outputs they can inspect: the interactive 3D scene and rendered presentation.

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 final published capture 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. The saved scene is also exported to GLB with geometry, materials, the contestant camera, and supported lights for interactive inspection. 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.

## Harness Instructions

You are operating Blender through the provided tools. The scene starts empty. Build what the user's prompt asks for in the current scene. Work step by step and verify your own work. Before you finish, set up a camera and lighting to frame a single, well-composed presentation image of your result — position and aim the camera, and add or adjust lights so the subject reads clearly. 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. The saved scene is also exported to GLB with geometry, materials, your camera, and supported lights for interactive inspection. When you are finished, reply with the single word DONE and no tool calls.
--TASK corgi-beach-ball: ## Task Complete the following Blender task: <task> A Pembroke Welsh Corgi caught mid-pounce on a beach ball, frozen at the funniest possible frame. This is a posed action still, not an animation — one moment, chosen well. The dog. Corgi anatomy is non-negotiable and is most of the test: a long low body, comically short legs, oversized upright triangular ears, a fox-like muzzle with an open pa…
16:38>>execute_blender_code({"code":"import bpy\nimport math\n\n# Clear scene\nbpy.ops.object.select_all(action='SELECT')\nbpy.ops.object.delete()\n\n# 1. Create the Corgi\n# Since we don't have a Corgi model, we'll use a combination of primitives to block out the anatomy\n# and then apply a subdivision surface for organic feel.\n# Long low body\nbpy.ops.mesh.primitive_uv_sphere_add(radius=0.4, location=(0, 0, 0.3))\nbody = …)
16:38<<Code executed successfully:
17:16>>get_viewport_screenshot({"user_prompt":"A Pembroke Welsh Corgi caught mid-pounce on a beach ball, frozen at the funniest possible frame. This is a posed action still, not an animation — one moment, chosen well.\n\nThe dog. Corgi anatomy is non-negotiable and is most of the test: a long low body, comically short legs, oversized upright triangular ears, a fox-like muzzle with an open panting mouth, a fluffy white-and-tan c…)
17:17<<
17:17imgimage delivered — get_viewport_screenshot
17:22>>execute_blender_code({"code":"import bpy\nimport math\n\n# The previous setup was too primitive. Let's refine the Corgi's anatomy and the ball's appearance.\n\n# 1. Refine the Corgi Anatomy\n# We'll use a more detailed approach: Sculpting using a single mesh with subdivision.\nbpy.ops.object.select_all(action='SELECT')\nbpy.ops.object.delete()\n\n# Create a base mesh for the corgi\nbpy.ops.mesh.primitive_uv_sphere_add…)
17:23<<Code executed successfully: Info: Applied modifier was not first, result may not be as expected
17:54>>execute_blender_code({"code":"import bpy\nimport math\n\n# The goal is to improve the Corgi's \"feel\" and the ball's detail.\n# We'll refine the colors, the pose, and add the \"tongue\" and \"ears back\" feeling.\n\n# 1. Refine Corgi Details\ncorgi = bpy.data.objects.get(\"Corgi\")\nif corgi:\n # Adding a tongue\n bpy.ops.mesh.primitive_uv_sphere_add(radius=0.05, location=(1.8, 0, 0.8))\n tongue = bpy.contex…)
17:54<<Error executing code: Communication error with Blender: Code execution error: 'bpy_prop_collection[key]: key "Transmission" not found'
18:05>>execute_blender_code({"code":"import bpy\nimport math\n\n# Fix the material error and refine again\nball = bpy.data.objects.get(\"BeachBall\")\nif ball:\n mat = ball.data.materials[0]\n nodes = mat.node_tree.nodes\n bsdf = nodes.get(\"Principled BSDF\")\n \n # In Blender 4.0+, 'Transmission' was replaced by 'Transmission Weight'\n # Let's check the input names\n inputs = bsdf.inputs\n transmiss…)
18:05<<Code executed successfully:
18:58>>execute_blender_code({"code":"import bpy\n\n# Ensure Cycles is used for the final result\nbpy.context.scene.render.engine = 'CYCLES'\n\n# Set up the camera to frame the corgi and ball properly\ncam = bpy.context.scene.camera\ncam.location = (3.5, -3.0, 0.6)\ncam.rotation_euler = (1.4, 0, -0.8) # radians\n\n# Set world background to a simple light blue sky\nbpy.context.scene.world.use_nodes = True\nbg_node = bpy.contex…)
18:59<<Code executed successfully:
19:07ai>DONE
19:07--status: ok · DONE declared
>
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