BlenderWorld: GPT-6-Astra Beats Opus-5 in 3D Benchmark with a Score of 79.9

The authors of the BlenderWorld benchmark propose testing not whether a multimodal model can write a Blender script, but how closely the finished scene matches the one that was requested. The set includes 223 tasks, and responses are scored against 2 651 binary rubrics plus an overall visual assessment. Among the thirteen frontier models tested, GPT-6-Astra posts the best combined result: 79.9. Opus-5 scored 73.0 and GPT-5.6-Sol 67.7. The authors' overall conclusion: models handle broad instructions but fail noticeably on fine-grained spatial consistency and accurate reproduction of structure.

How the test works

Multimodal models increasingly work as agents: they write code for a 3D editor or control the program directly. A real task rarely sounds like "make a cube." Usually it is a set of references: a photo of a building, a video of a product, a diagram with dimensions, a description of the lighting. From these, the model has to infer shape, materials, object placement and light without violating the instructions. According to the authors, existing benchmarks cover such open-ended requests poorly.

BlenderWorld fills this gap across four areas: architecture, products, natural environments and scientific visualization. The inputs vary too: text, images, video and structured specifications. The tasks require reconstructing an object, modifying an existing one or designing a new one.

Scoring is the same for everyone. The final Blender projects are rendered from several angles, and the output is checked against rubrics written for each task. This matters: a single good frame can hide a geometry error that is visible only from the side. An overall visual score is calculated separately, and the final fused score combines the two approaches.

• Tasks: 223
• Binary rubrics: 2 651, about 12 per task
• Models compared: 13
• Input formats: text, image, video, structured specifications
• Leaders: GPT-6-Astra — 79.9; Opus-5 — 73.0; GPT-5.6-Sol — 67.7

Where models break down

The gap between first and second place is 6.9 points, and between second and third it is 5.3. From the leader to the third-place model the difference adds up to 12.2 points, so the choice of model clearly affects the result. But even the best score is below 80, which means the winner still fails a significant share of the binary checks.

The authors state the cause specifically. Models read the task and identify the right objects, but are worse at maintaining the relative position of parts, proportions and structure. These errors are harder to spot than a wrong color: the wall is in place and the window is there, but it is shifted, and the staircase does not reach the floor above. Binary rubrics applied to multiple camera angles exist for exactly this reason, since they check each detail separately.

The description includes no detailed breakdown by domain or input type, so it is not yet possible to say which is harder, video or specifications. There is also no data on how many attempts and tools each agent had. These parameters usually have a large effect on the numbers, and they are worth checking in the full version of the paper.

What teams should do

For studios and product teams considering 3D automation, the takeaway is practical. Agents can already be used for draft scenes, blocking out and generating variations, where overall composition and style matter. Final geometry, dimensions and joints will still have to be checked by a person or by separate automated tests.

For your own evaluations, it is worth adopting the BlenderWorld approach: render the result from several cameras and check it against a list of binary criteria, rather than relying on a single impression from a preview. Rubrics such as "the table legs stand on the floor" or "the windows are symmetrical relative to the door" are cheap to verify and catch exactly the errors that models stumble on.

Finally, the benchmark's numbers point to potential for models built around 3D from the start: the authors themselves speak of substantial headroom for future 3D-native agents. For now, the results are useful as a guide when choosing a model and as an honest list of weak spots to account for in your pipeline.