Glonce

Simon Willison drives Blender with ChatGPT Codex on macOS

toolsThe Glonce editorsSimon Willison🔥43Glonce rating43Impact30Novelty25Relevance80Surprise20Players35Credibility75
Simon Willison drives Blender with ChatGPT Codex on macOS

Simon Willison published a TIL post on 5 September 2026 describing how he has been using ChatGPT Codex, a coding agent, to control Blender on his Mac. He installed the full Blender desktop application from blender.org, then gave the agent a prompt telling it to use the already-installed copy at /Applications/Blender to render a scene of a pelican riding a bicycle. He followed up with two more prompts in the same session: first asking it to add a background and a lot of flair, then to make the result a whole lot better. The final image was generated through Blender's Python API, which the agent used to build and manipulate the 3D scene. Willison notes more broadly that modern frontier models have gotten really good at using Blender this way: beyond single rendered images, the same approach can produce editable .blend project files and even full movies, made by rendering a sequence of frames and stitching them together afterward with ffmpeg.

Key facts

  • Simon Willison published the TIL post on 5 September 2026, describing his own recent experiments.
  • The coding agent used was ChatGPT Codex, running on macOS against a locally installed copy of Blender.
  • Setup: install the full Blender Mac app from blender.org, then prompt the agent to use the already-installed /Applications/Blender.
  • The example task started with 'render a scene of a pelican riding a bicycle', then iterated with 'add a background and a lot of flair' and 'make it a whole lot better'.
  • The final image came from Blender's Python API; the same technique can also output editable .blend files and full movies stitched from rendered frame sequences via ffmpeg.

Why it matters

It is a concrete data point that frontier models paired with a coding agent can now drive a full 3D content creation application, not just write and run code. Willison frames it as models having gotten really good at using Blender specifically, turning natural-language prompts into working .blend scenes, rendered images, and even movies assembled from image sequences with ffmpeg.

Who it affects

Developers and hobbyists who already use coding agents like ChatGPT Codex and want to extend them into 3D graphics and rendering work on a Mac, without learning Blender's interface or Python API directly.

How to use it

Install the full Blender desktop application from blender.org. Then prompt a coding agent such as ChatGPT Codex to use the already-installed /Applications/Blender copy for a specific rendering task, and refine the result with plain-language follow-up prompts, as Willison did by asking for a background and flair and then a general improvement pass.

How solid is it

This is a single first-person TIL account from one author using one coding agent (ChatGPT Codex) on one platform (macOS), not a benchmark or a comparison across models or tools. It documents that the workflow worked for this specific pelican-on-a-bicycle example, not how reliably it generalizes.

Risks and caveats

The post gives no figures for rendering time, image resolution, cost, or hardware used, and does not say how many iterations a typical task needs beyond the two follow-up prompts shown. It also does not say whether the technique works with coding agents other than ChatGPT Codex.

“Modern frontier models have got really good at using Blender.”

— Simon Willison