Astra performance very poor
What version of the Codex App are you using (From “About Codex” dialog)?
26.917.62051
What subscription do you have?
Pro
What platform is your computer?
Microsoft Windows NT 10.0.26200.0 x64
What issue are you seeing?
I used Astra High all weekend to create a vehicle for a simulation game. I gave it detailed instructions and highly detailed concept art that clearly defined the vehicle from every angle. I even pointed Astra at many example vehicles in the game. I gave Astra permission to use Computer Use skill to test in game , provided detailed instructions on how to test in game, and even created a virtual gamepad for it to help it test.
The result was low poly, low detail, junk. Astra burned a significant amount of my 7d usage and produced nothing usable. There were pieces of the vehicle just floating in mid air. It looked like something you'd see in Minecraft. It looked nothing like the concept art I provided. The vehicle interior looked like a three year old kid drew it. It's like Astra designed the vehicle for a lego game. There is nothing realistic about the vehicle it created.
This was very disappointing compared to work that I've had GPT-5.6 Sol High produce. I thought Astra was suppose to be better? Astra's reasoning was lacking, it's performance slow, and it's Computer Use and vision skills must be much worse to produce what it did. With this result, it's unclear why Astra burns my usage much faster than GPT-5.6 Sol High did.
I told Astra I was disappointed with the result and gave it screenshots of all the issues. It's answer was:
You’re right. Candidate G fails your visual requirement. Calling its appearance “unaccepted” was too soft—it should have been marked FAIL.
Your screenshots show clear problems: floating trim and light strips, disconnected overlapping panels, visibly faceted parts, crude mechanical assemblies, and a cockpit built from basic shapes without convincing construction or detail. The inverted monitor and obstructed instruments make that worse. This does not meet the standard established by your concept art and ... references.
I made the wrong decisions. I kept adding small details to a fundamentally inadequate model, then spent substantial usage rebuilding, packaging, and testing it. Correct dimensions, working lights, and passing ... checks could not establish the visual quality you requested. I should have rejected the modeling approach much earlier, based on the screenshots I already had.
I also should not have shifted that judgment back to you as “visual acceptance outstanding.” Recognizing those obvious defects was my responsibility.
A credible recovery requires rebuilding the main body surfaces and cockpit, with convincing connected construction—not another collection of cosmetic additions. I have not demonstrated that this workflow can deliver your requested standard, and I cannot honestly promise that another automated pass will achieve it.
I’m sorry for wasting your time and paid usage. I won’t start another modeling or testing run in response to this complaint; the project files and installed mod remain untouched by this response."
I now have to start all over with GPT-6 Sol and hope for the best. Unfortunately I only have 7% left of my 7d and it doesn't reset until Oct 3rd...
What steps can reproduce the bug?
Give Astra concept art, demonstrated Blender usage skills, demonstrated Computer Use instructions, and demonstrated native game testing instructions. Ask it to produce the 3d model from the concept art.
What is the expected behavior?
A result that is somewhat close to the concept art.
Additional information
Sessions:
- 01a0e2c6-1c7b-71f3-9dfd-40edc788f332
- 01a0dff6-fde6-7b83-860b-3cd455ae47b4
- 01a0df13-01f5-71a3-8814-4db8efb7670f
- 01a0ded0-437e-7233-83e1-7bdafb8c9272
- 01a0e7d9-aa8e-75f1-bc29-80d5027c69f9
Source: openai/codex