Two images, two editable Blender studios

Tested November 17, 2026 on the same Mac. Astra completed the build and edit in 11m 20s, compared with 17m 52s for Sol. Both scenes opened on their own, rendered a previously unseen view, and supported moving the whole chair and changing the lighting without operator repairs.

Original studio reference on the left, Astra reconstruction in the centre and Sol reconstruction on the right

Test conditions

The reference is an original synthetic Blender studio. Fresh signed-in Codex sessions received the same two PNGs in the same order and the same build prompt. The sessions were configured for GPT-6.2 Astra and GPT-6.2 Sol at High effort and default speed. Astra ran first. Each model had a separate temporary working directory. The third image, reference generator and scoring data stayed with the evaluator. Both builds were archived before the sessions received C’s camera coordinates. Neither session received C.png. The command logs contain no reference to an evaluator file. Blender 5.2.1 LTS rendered all scored frames at 1280 × 800 with Cycles CPU, 24 samples and denoising. CPU thread counts were left to the models for their own renders. The independent checks used six threads.

Time, usage and repairs

Measurement Astra Sol
First observed scene save, elapsed from build prompt 4m 07s 13m 51s
Complete build submission 9m 39s 15m 37s
Camera, chair and lighting edit submission 1m 41s 2m 16s
Total active model-session elapsed time 11m 20s 17m 52s
Render time stamped on the five delivered PNGs, combined 27.05s 29.47s
Follow-up repair rounds 0 0
Manual scene fixes and hands-on repair minutes 0 / 0 min 0 / 0 min
Combined input tokens, including cached reads 1,839,343 2,804,624
Cached input tokens included above 1,750,016 2,696,320
Combined output tokens, including reasoning 18,437 31,457

Phase times include model iteration, tool calls, rendering and startup retries. Totals sum the unrounded session times and exclude waiting for the other model and independent QA. First save reflects the first observed project on disk. The completed submissions were independently reopened. Final PNG stamps exclude preview renders and verification reruns. Usage comes from the final cumulative session counters. Both sessions used the existing signed-in Codex plan with API-key variables removed. Per-run dollar charges and subscription credit allocations were unavailable.

Both models encountered a sandboxed Blender startup crash and recovered during the timed build. Astra also switched Python environments after an image-library import failed. Astra refined the monitor, keyboard and wall slats. Sol iterated on lighting, wood tones and the print. These were model-led changes within their first submissions.

Physical dimensions

Centimetres, measured from the saved desktop mesh in world coordinates. The last number is the height of the top surface above the floor.

Dimension Reference Astra Absolute error Sol Absolute error
Width 165.0 170.0 5.0 162.82 2.18
Depth 80.0 80.8 0.8 76.67 3.33
Top height 78.0 77.925 0.075 74.5 3.5

Astra assumed a 170 cm desk width and came close on height and depth. Sol assumed a 74.5 cm desk height and came closer on width. Neither model received a known scale measurement.

Hidden-view accuracy

Manual pixel coordinates use the top-left as origin. Five reference marks were recorded before the runs. Both renders showed all five. All three images use the full 1280 × 800 frame without alignment, cropping or resizing for measurement. Rounding is to the nearest pixel for distances.

Landmark Reference (x, y) Astra (x, y) Error, px Sol (x, y) Error, px
Monitor upper-left corner 481, 313 483, 306 7 480, 324 11
Front-right desktop corner 725, 570 743, 571 18 717, 578 11
Microphone lower tip 684, 377 697, 371 14 691, 390 15
Poster upper-left corner 730, 70 730, 62 8 728, 82 12
Globe centre 1248, 364 1268, 362 20 1239, 374 13
Median across five 14 12

The two-pixel difference between medians is within the few-pixel precision of marking rounded edges and a partly cropped globe by eye. The most visible differences are the desktop extent, the outline of the microphone arm, the contrast between the wall slats and the amount of the shelf globe in frame. Both models recovered the relationship between the desk, chair, poster and microphone. This was one run per model on a clean synthetic room.

Edit and portability checks

Fresh Blender processes opened each original .blend in a new directory containing only that file. Both rendered camera A with zero decoded pixel differences from the submitted PNG. Both original files had no image assets or missing textures. Each file retained independently editable mesh objects, materials and lights.

From each original file, the evaluator recreated C's camera, rendered it, moved the chair parent 0.30 m in positive X, and rendered again. Astra’s two replays matched the submitted PNGs exactly. For Sol each replay differed by less than 0.006 intensity levels per channel on average on a 0–255 scale. In both warm files the chair and other original geometry returned to their original positions. The Key light moved exactly 1.00 m in positive X and its linear RGB was (1.00, 0.55, 0.28). Other lights retained their settings and Key retained its power. The original build files remained byte-identical in the edit submissions.

Input A · Input B · Held-out reference · Astra render · Sol render

Astra edits · Sol edits