How to Use GPT-6 Astra in Blender with Codex

To control Blender through Codex with GPT-6 Astra, you do not need to master every 3D modeling tool. Simply describe the desired end result, and Codex will choose an available approach: driving the interface via the official Computer Use plugin, running Blender Python (bpy) scripts, or executing background Blender commands. You do not need to decide which tool the agent should apply at every step. It is enough to review intermediate results and save your project at each milestone.

Choose one of two ways to connect the model:

Your Situation	How to Connect
You have an account and a subscription with access to Codex and Astra	Sign in to your official account and select GPT-6 Astra
You do not have a subscription	Connect a BetterToken API Key; no official Codex account sign-in is required

In both cases, install the exact same official Computer Use through the Plugins section. The API Key changes how you access the model and pay for requests; the plugin installation process remains identical.

Below is the complete walkthrough from software setup to an editable desk lamp and a turntable orbit animation. You will need a Mac or Windows computer capable of running Blender. The model receives tasks over the internet, while Blender and rendering run locally on your machine.

In the official OpenAI write-up, Astra creates a more complex architectural scene. This guide uses a compact desk lamp project completed in the environment described below.

1. Install the App and Set Up Model Access in Codex

Visit the official ChatGPT download page, select the desktop application for your operating system, and install it. In the current application, Codex is located in the mode switcher. If you already have a standalone Codex App installed, update it and use the same Codex section.

If You Have a Subscription

1. Launch the application and select sign in with ChatGPT.
2. Sign in to the account where your subscription is active.
3. Switch to Codex.
4. In the model selector next to the prompt input, find GPT-6 Astra. If it is not listed, verify the models available to your account.

A separate BetterToken Key is not required for this method. Proceed directly to creating the project.

If You Do Not Have a Subscription: Connect a BetterToken API Key

This option also works without an official Codex account. You use your BetterToken account, and model requests are billed based on usage.

Create a BetterToken API Key to use GPT-6 Astra in Codex for working with Blender.

If Codex Already Works via an API Key

If you already have an active API connection, simply replace the Base URL and API Key with your BetterToken credentials. There is no need to run the installer script again.

1. In BetterToken, create a key for the GPT provider.
2. Open your current Codex API provider settings. Replace the Base URL (base_url) with https://www.bettertoken.ai/v1, and replace the current API Key with your BetterToken key. If the key is stored in an environment variable, update its value.
3. Ensure that gpt-6-astra is selected as the model. If a different model was configured previously, update its identifier.
4. Save the settings, completely restart Codex, and send a short test message to verify the connection.

Once the test succeeds, proceed to setting up the project and Blender. If Computer Use is already installed, you can continue using it; if not, install it following the steps below.

If You Are Configuring an API Connection for the First Time

1. Open your Workspace in BetterToken and create an API Key for the GPT provider. Copy the key to configure it on your machine.
2. In the key connection setup, select Codex and the gpt-6-astra model. The full identifier is important: do not shorten it to astra.
3. Open the BetterToken Codex guide and navigate to the Command-line setup section. The installer requires Node.js; if it is not yet installed, install it first according to the setup instructions.
4. Close Codex completely. Open Terminal on macOS or PowerShell on Windows, then run the corresponding command below for your operating system.
5. When the installer prompts for your API Key, paste your key locally. Do not send it into the chat with the agent. Wait for the configuration and connection test to complete.

macOS — Terminal:

curl -fsSL "https://www.bettertoken.ai/install-codex-provider.sh" -o /tmp/install-codex-provider.sh
bash /tmp/install-codex-provider.sh --model gpt-6-astra

Windows — PowerShell:

iwr "https://www.bettertoken.ai/install-codex-provider.ps1" -OutFile "$env:TEMP\install-codex-provider.ps1"
powershell -ExecutionPolicy Bypass -File "$env:TEMP\install-codex-provider.ps1" -Model "gpt-6-astra"

The installer configures the BetterToken provider and connects via the Responses API, then performs a quick model check. Note that the verification request also consumes API tokens. Once finished, reopen the application and switch to Codex. Signing in via ChatGPT is not required for this workflow.

If the configuration fails with an error, resolve it first: verify your key, check access to the model, and ensure your account has sufficient balance for API calls. Installing Computer Use will not fix a model connection failure.

2. Create a Local Project

1. Create an astra-blender folder on your computer—for example, inside Documents.
2. In Codex, click ”+” next to Projects and add this folder as a local project. The button label may vary depending on your application language and version; choose the option to add a project on your computer.
3. Open a new task inside it and confirm that Astra is selected.
4. Send a short test message: “Reply with one word: ready”.

Receiving a response without connection errors means you are ready to proceed to Blender. For API connections, a record of this call should also appear in your BetterToken usage history. This verifies the model connection; access to the Blender window will be tested separately.

The project folder will store your .blend file, rendered images, and animation files. Return to this same task for subsequent edits so Codex retains context from earlier instructions.

3. Install and Open Blender

Blender is a free 3D creation suite covering modeling, materials, animation, and rendering. You can download it from the official Blender download page.

1. Select the installer for your operating system. On Mac, make sure to choose the correct architecture: Apple Silicon or Intel.
2. On macOS, open the downloaded disk image and drag Blender into Applications. On Windows, run the installer and follow the setup wizard.
3. Launch Blender. On first startup, keep the default settings and select a new General file.
4. By default, you should see the default cube in the center, and the scene outliner on the right showing Cube, Camera, and Light. Do not delete them yet.
5. Go to File → Save As, navigate to the astra-blender folder, and save the file as desk-lamp.blend.

If you prefer to let Codex handle the installation, send this prompt inside your created project:

Help me install Blender from https://www.blender.org/download/ on this computer.
First, check whether Blender is already installed. Use the official stable build.
Explain any installer steps or system permissions that I need to complete manually.
Launch Blender and confirm which version is running.

The installation is complete when the Blender window itself is open. A download link in Codex’s response is not sufficient. Official installation guides for all supported operating systems can be found in the Blender Manual.

4. Install the Official Computer Use Plugin

The remaining steps are identical whether using a subscription or a BetterToken API Key:

1. Inside the application, navigate to Plugins.
2. Locate Computer Use by OpenAI and click Install plugin.
3. If the button displays Enable after installation, turn the plugin on.
4. On the plugin card, enable both its MCP server and Skill—the server provides the control tools, while the Skill supplies instructions on how to use them.
5. Go to Settings → Computer use and enable Any App application control if that toggle is present in your version. Grant permission to control Blender when prompted by the app.
6. On Mac, grant the requested Screen Recording and Accessibility permissions in System Settings. Restart the application if prompted by macOS.
7. Return to your project task. If necessary, open a new task so the newly installed plugin tools load properly.

The official Computer Use documentation illustrates the exact layout of these settings. On Windows, keep Blender visible on an active desktop and do not lock your computer while the agent is running: it interacts with the active session’s mouse and keyboard.

Now test direct interface control:

Use the official Computer Use plugin to inspect the open Blender window.
Tell me the current filename and the objects visible in the scene list.
Using the Blender interface, rename Cube to Connection_Test.
Show the updated scene list, then stop. Do not create or delete any objects.

Check the Blender window: the cube should remain in place, and its name in the scene list should change to Connection_Test. If Codex merely describes what to do without executing, explicitly instruct it to perform the actions via Computer Use. This small check confirms that the agent can see and interact with the target window.

5. Build Your First Model

Start with an object that is straightforward to verify: a desk lamp with a circular base, a stem, an angled arm, and a shade. Hand the full task to the agent and allow it to choose the most reliable available tool.

Create a small desk lamp in the current Blender project. If necessary,
you can use Computer Use, Blender Python (bpy), or Blender background
commands. Choose the most reliable available tool and do not assume
that Blender MCP is installed.

Replace only the Connection_Test object from our connection check.
Create a circular base resting on a flat surface at Z=0, a vertical
stem, a short angled arm, a conical lampshade with wall thickness,
and an internal lightbulb. Keep the parts as separate, editable objects
named Lamp_Base, Lamp_Stem, Lamp_Arm, Lamp_Shade, Lamp_Bulb, and Ground.

Verify the actual result for alignment, proportions, and part connections.
Set up a camera that shows the entire lamp, save desk-lamp.blend in this
project, and show front and perspective views before stopping for feedback.

How the Real Run Went

This real run took place on macOS in ChatGPT desktop app version 26.908.40834, using Blender 5.2.1 LTS and the GPT-6 Astra High model. First, Computer Use renamed the cube to Connection_Test. Codex then authored bpy scripts and executed them using Blender’s built-in Python. The user did not need to install a separate Python environment, configure virtual environments, or micro-manage tool selection.

When keyboard input through the open Blender interactive console became unstable, Codex switched to the Blender background CLI, finalized the connection between Lamp_Arm and Lamp_Shade, and prepared the renders. After that, the agent reopened the saved .blend file in the Blender GUI to inspect the result visually.

The scene retained separate objects named Lamp_Base, Lamp_Stem, Lamp_Arm, Lamp_Shade, Lamp_Bulb, and Ground. The bottom of the base was confirmed at Z=0, the lampshade maintained proper wall thickness, and the bulb was seated inside the shade.

Check these same details in your own scene: the base rests flat on the ground plane, the stem connects cleanly to the shade, the entire lamp fits within the camera frame, and parts can be selected individually. Correct one issue at a time and save the file after each successful adjustment.

6. Add Materials and Lighting

Once you are satisfied with the geometry, move on to materials and scene appearance:

Continue working with the existing desk lamp in desk-lamp.blend. Keep its
geometry and object names intact. Assign a dark green enamel finish to Lamp_Base
and Lamp_Shade, brushed brass to Lamp_Stem and Lamp_Arm, and a warm glow
to Lamp_Bulb. Add a neutral material to Ground and set up three-point studio
lighting.

Use Cycles to render a clean preview from the camera. Inspect the image
for floating parts, intersections, framing issues, or abnormal reflections,
then save the preview and desk-lamp.blend.

In our run, the base and shade received a dark green enamel finish, the stem and arm were assigned brushed brass, and the bulb was given a warm emission glow. Codex set up three-point studio lighting and rendered the result in Cycles. The final render is shown at the beginning of this article.

For Further Work: Connect the Official Blender MCP

Computer Use operates the visual interface. MCP provides Codex with a dedicated communication channel to a running Blender instance, allowing it to inspect scene data and execute operations directly. In the workflow described above, Blender MCP was neither installed nor required: Codex relied on Computer Use and Blender’s built-in Python/CLI capabilities. MCP remains an optional enhancement for future projects.

1. Visit the Blender MCP page and proceed to the official Setup guide.
2. Check your Blender version: the official project currently requires 5.1 or newer.
3. In the same Codex task, provide the Setup link and ask it to set up both the Blender add-on and the standalone MCP server, then register the server in Codex. These are two distinct components of the connection.
4. Once the add-on package is ready, open Blender and go to Edit → Preferences → Add-ons → Install from Disk, select the ZIP archive, and enable the add-on. Codex can guide you through this step using Computer Use.
5. In the add-on preferences, start the connection server. The add-on and the MCP server must share the exact same host address and port; use the values specified in the current official guide.
6. Open a new task in the same Codex project and ask it to fetch the object list via MCP. Confirm that Lamp_Base, Lamp_Shade, and the other object names match Blender.

Do not proceed with modifications until the object list matches perfectly. From there, you can delegate geometry building to MCP while using Computer Use to inspect the window and verify rendered frames. If the connection drops, always inspect the current scene state first to avoid creating duplicate objects.

7. Create a Short Animation

Once you have saved a solid static model, instruct Codex to set up a 360-degree turntable camera animation. You can keep the animation in the same desk-lamp.blend file:

In the existing desk-lamp.blend scene, create a 5-second 360-degree camera
orbit at 24 fps. Set keyframes at frame 1 (0 degrees) and frame 121 (360 degrees),
and render frames 1 through 120 so the loop does not duplicate the endpoint.
Use Cycles at 1280x720 with 32 samples. Keep the lamp stationary and all parts
editable.

First, render test frames 1, 41, and 81. Once confirmed, render all 120 PNG
frames into the renders/ folder and encode a 5-second H.264 MP4 named
desk-lamp-orbit.mp4. Save the scene in desk-lamp.blend.

In the completed project, Codex first rendered test frames 1, 41, and 81. Cycles with 32 samples then rendered 120 PNG frames at 1280×720 into the renders/ directory. Frame 121 held the 360° position but was excluded from the final output, ensuring that the start of the 5-second loop is not duplicated at the end.

The system FFmpeg binary failed to execute due to a missing libvpx.11.dylib library. Codex checked the capabilities of Blender 5.2.1 and, rather than troubleshooting the system installation, leveraged Blender’s built-in H.264 video encoder. Before compiling, it verified that all 120 PNG files existed and had non-zero file sizes, then configured the H.264 MP4 export for 120 frames at 24 fps and 1280×720. After encoding completes, open the video file and verify that it plays back smoothly for the full five seconds.

If frame rendering is slow, ask Codex to report the progress of the active render. Never launch a second render of the same scene until you confirm whether the first one has finished.

Troubleshooting

Codex doesn’t open or control Blender

Work through these checks in order:

1. Computer Use tools are missing: check the plugin installation, turn on both the server and Skill toggles, and start a new task.
2. Tools are present, but the Blender window is inaccessible: grant access to Blender; on Mac, verify Screen Recording and Accessibility, and on Windows, keep the active desktop unlocked.
3. The agent only describes the steps: explicitly ask it to perform one small action via Computer Use and show the state afterward.

Other symptoms

What You See	What to Check
Codex does not respond, reports 401, or cannot find the model	For an API connection, verify your key, the GPT provider, gpt-6-astra, and the setup script output; completely restart the application
The model responds, but Computer Use tools are missing	Check the plugin installation, verify both the server and Skill toggles, and start a new task
Tools are present, but the window is inaccessible	Grant permission to access Blender; on Mac, verify Screen Recording and Accessibility, and on Windows, ensure the active desktop is unlocked
The agent describes the steps but does not change anything	Explicitly ask it to perform a small action via Computer Use and show the state afterward
MCP is connected, but shows a different scene	Verify the currently open .blend file, the active Blender instance, and the extension address/port
The call was interrupted or timed out	First request the list of objects and saved files; resume from the last completed stage
The system FFmpeg fails to launch	In this run, Codex used Blender’s built-in H.264 encoder; this is a reliable workaround if the current Blender build supports the required format
The image looks poor	Step back to basic geometry and a single camera angle; fix proportions, intersections, and lighting before attempting a full render

After completing this exercise, you should have desk-lamp.blend, verification images, 120 PNG files in the renders/ folder, and desk-lamp-orbit.mp4. Open .blend once more and give Codex one specific edit to confirm that the parts remain individually editable.