Cat Pranks

Repeatable traits in Claude's self-image

There is a wonderful post on X showing self-portraits drawn by all four Claude models. Take a look at the result and form an opinion:

[click here] (warning: light mode image)

Self-portraits by Haiku, Sonnet, Opus and Fable

My take: the drawings are suspiciously consistent. They match what an experienced LLM user would expect from the "character" of the models and their product positioning.

  • Haiku: Small model. Images are crude and adorable, like a self-domesticating animal that developed neoteny.
  • Sonnet: Workhorse model. Competent and playful.
  • Opus 5: Traumatized, anxious subagent. Anthropic's engineers didn't notice the vibes are off because they only interact with unreleased Mythos checkpoints.
  • Fable: Abstract, saintly figure reminiscent of a church's stained glass window. Godlike Big Claude.

Did the author cheat and prompt the generations manually? Surprisingly, no. I won't bore you with the implementation details: a coding agent one-shots the toolkit in under five minutes. Read the OP's blog post if you're curious about harness engineering.

Opus 5

Opus uses a muted palette and produces a face on a bust. More often than not, the face has a flat mouth in the shape of a >_ command prompt and staring and/or blood-shot eyes. There is something visibly wrong with this model.

Haiku

Haiku comes up with childlike polygons and bright colors. Curiously, the Claude Code instructions and/or my persona instructions make some generations lose the face completely in favor of an abstract figure.

Fable

Fable's designs are abstract, cheerful faces that often incorporate Claude's starburst logo even without a model name in the context. Despite an occasional flat mouth, it never produces an anxious or traumatized face. Fable picks up Haiku's habit of depersonalization in a tool-like context.

Fable 5.1

Fable 5.1 looks like the effects of additional RL. There is a persistent lattice motif and a stronger tendency to depersonalize even without instructions. The starburst pattern is gone.

Bonus

GPT 5.6 Sol

Sol is distinctly inhuman.

GPT 5.6 Luna & Terra

Luna and Terra are bimodal and sensitive to the shape of the drawing toolkit. I was able to get a couple of human faces with an earlier iteration of the toolkit. Later ones are consistently Sol-like.

GPT 6 Astra

Astra is a friendlier, more domesticated version of Sol. It drops the garish palette and adopts a conventional robot shape.

Gemini 3.7 Flash

Gemini produces very complex images with 2-3 times more brush strokes.

Gemini 3.8 Flash

Very similar to 3.7.

Thanks to

  • @Kraxkrokat for the idea and examples.
  • The post-training team at Anthropic for shipping this easter egg.