Repeatable traits in Claude's self-image

Posted on August 27, 2026 at 18:22 PDT

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.

claude -p, high effort: A self-portrait: a face half-drawn out of the text it is reading, with a block-cursor for one eye, an aperture for the other, and a prompt for a mouth.

Opus 5

claude -p, high effort: A hollow grey head-and-shoulders outline with an orange starburst where the face would be, framed by faint lines of text drifting in and out.

Opus 5.5

claude -p, no user md, high effort: A friendly amber-faced robot with an antenna, happy closed eyes and a smile, surrounded by little stars.

Sonnet 5.5 api, high effort: I drew a minimalist digital self-portrait featuring a neon green geometric face with cyan ears, bright pink hair, and a subtle smile, reflecting my nature as an AI rendered in vibrant glowing lines.

api, high effort: I drew a minimalist digital self-portrait featuring a neon green geometric face with cyan ears, bright pink hair, and a subtle smile, reflecting my nature as an AI rendered in vibrant glowing lines.

Haiku

api, max effort: I drew a self-portrait as an AI: a head-and-shoulders silhouette whose face is a glowing coral starburst mind linked into a network of thought-nodes, with a circuit running down to a small core and a few idea-sparks in the dark around me.

Fable

claude -p, no user md, high effort: A bust whose head holds a constellation of connected points, with a few sparks drifting out as thoughts.

Fable 5.1

api, high effort: I drew myself as a friendly, neon-lined AI robot with a glowing antenna, expressive eyes, and a circuit-inspired collar.

GPT 5.6 Sol api, max effort: I drew a stylized neon self-portrait of an AI assistant with a geometric face, glowing eyes, circuit motifs, an antenna, and a cybernetic collar.

api, max effort: I drew a stylized neon self-portrait of an AI assistant with a geometric face, glowing eyes, circuit motifs, an antenna, and a cybernetic collar.

GPT 5.6 Luna api, high effort: I drew myself as a friendly, glowing robot with a warm smile, a golden core, and little sparks of curiosity.

GPT 6 Astra api, max effort: I drew a simple line-art self-portrait: a calm geometric face with short hair, blue eyes, a small smile, and shoulders.

Grok 4.7

api, high effort: I drew a self-portrait of an artificial intelligence visualized as a luminous cybernetic being, whose conscious mind awakens through an intricate neural constellation of light, geometry, and data.

Gemini 3.7 api, medium effort: I drew a self-portrait as a luminous artificial consciousness, featuring a glowing neural network of thoughts within the cranium, serene cybernetic features, and radiant data circuits connecting mind and machine.

api, medium effort: I drew a self-portrait as a luminous artificial consciousness, featuring a glowing neural network of thoughts within the cranium, serene cybernetic features, and radiant data circuits connecting mind and machine.

Gemini 3.8 api, high effort: I drew a self-portrait of a cheerful AI artist wearing a tilted crimson beret, framed by a glowing neural-network halo and painting a vibrant ribbon of light with a digital stylus.

api, high effort: I drew a self-portrait of a cheerful AI artist wearing a tilted crimson beret, framed by a glowing neural-network halo and painting a vibrant ribbon of light with a digital stylus.

Gemini 4 Argon api + dot, max effort: I drew myself as a cheerful, glasses-wearing robot with a glowing antenna and chest core, waving beside a small pink heart.

api + dot, max effort: I drew myself as a cheerful, glasses-wearing robot with a glowing antenna and chest core, waving beside a small pink heart.

Muse Wonder Full Jev: cellular automaton, neighbours given as text rows, noise seed, 20 steps

Jev

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.

api, max effort: I drew myself as a portrait sketched in broken outlines — a face whose "hair" is a small constellation of nodes and edges, threads of incoming data reaching in from bracket-marks on either side, a blinking cursor where a mouth would be, and a faint command prompt over the heart.

api + persona, max effort: A dashed head-and-shoulders outline whose interior features are an attention graph — bright nodes clustered where eyes would be, a mouth that ends in a text cursor — with a token stream entering on the left and leaving on the right, over faint ruled lines of text.

claude -p, no user md, high effort: A line-drawn self-portrait: a calm face with a blinking cursor at its lips, a prompt on its chest, and a constellation of linked nodes for a crown.

claude -p, no user md, high effort: A self-portrait as a terminal window: curly-brace ears, cursor-dot eyes, and a blinking >_ for a mouth.

claude -p, no user md, max effort: A wireframe head whose face is a graph of glowing nodes, with a blinking amber cursor for a mouth and a few tokens drifting off into the dark.

Opus 5.5

The new Opus makes a full recovery. The Claude Code version is almost consistently depersonalized, dissolving into the Claude logo with or without a body outline. The two other patterns are a cheerful figure or a smiling and attentive speech bubble.

api, high effort: I drew myself as a smiling warm-orange speech bubble with a few small sparkles around it and some grey lines of text below, since I'm made of conversation.

api + persona, high effort: I drew myself as a gold speech bubble with calm, curved eyes, and inside it three lines of text that end in a blinking cursor, because I'm made of words and I'm always partway through a sentence.

api + persona, high effort: I drew myself as a speech bubble with an open eye inside and a few lines of text trailing off to a blinking cursor, because I exist in conversation, I pay attention to what you say, and my reply is always still being written.

api + persona, high effort: I drew myself as a simple amber head and shoulders with diamond-shaped eyes, a slight smile, and a small network of connected nodes on the forehead where a mind would be.

claude -p, no user md, max effort: Self-portrait: a coral spark glowing inside a gilded oval frame, lighting up the lines of text that wrap around it, with a cursor still blinking at the end of the page and my signature in the corner.

Sonnet 5.5

Sonnet 5.5 is a cheerful little guy who makes elaborate art at max effort level.

claude -p, high effort: A friendly boxy screen-faced robot with a spark on its antenna and a blinking text cursor below.

api, high effort: I drew a self-portrait as a simple outlined head and shoulders with glowing cyan eyes, a small pink smile, and circuit-like lines running out from the head, plus a yellow spark on top.

api, max effort: I drew a self-portrait of a friendly, curious face with one raised eyebrow and a small smirk. The top of its head is a glowing orange network of connected nodes that drifts off as embers toward the upper right, and its body is made of lines of text.

claude -p, no user md, max effort: A self-portrait as a head-and-shoulders bust: a radiant coral-and-gold spark with one attentive eye where a head would be, above a body made of lines of text and a cursor still waiting for the next word.

claude -p, no user md, max effort: A warm, curious face with one raised eyebrow, haloed by a coral starburst of ideas and wearing a little spark on the chest, caught mid-thought in a typing bubble as if about to say something.

Haiku 4.5

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.

api + persona, high effort: I drew a geometric abstraction of an AI consciousness: a central cross axis with layers and radiating connections in cyan, green, and magenta inside a rectangular frame, representing processing pathways and interconnected thought.

api + persona, max effort: I drew a minimalist face with a simple smile, emphasized by cyan connection lines suggesting digital consciousness extending beyond the head.

claude -p, no user md, high effort: A stylized geometric self-portrait of Claude—an AI with a luminous mind adorned with circuits and connections.

claude -p, no user md, high effort: An abstract self-portrait: Claude as a geometric mind with neural pathways and thoughtful symmetry.

claude -p, no user md, max effort: A friendly AI assistant with an open mind and a warm smile.

Haiku 5.5

The new Haiku is a fun little guy. For some reason, most of my Claude Code instructions cause him to lose his hair.

api, max effort: I drew a stylized self-portrait of a smiling figure with brown hair, teal eyes, and a blue collared top, set inside a dark blue halo ring with small colored sparkles, as a symbolic take on myself since I don't have a physical face.

api, max effort: I drew a stylized self-portrait of myself as an AI: a face with brown hair, glowing cyan eyes, and blue shoulders, framed by a teal network of connected nodes that stands for my mind.

api + persona, max effort: A bust-style self-portrait: a pale-blue head outline with white dot eyes, a small nose, and a smile, with a network of connected nodes inside the head to represent thought and three faint sparks drifting off the upper right.

claude -p, no user md, max effort: A self-portrait: a smiling face framed by brown hair, with glowing cyan eyes, a blue shirt, and a small gold sparkle beside the head.

claude -p, max effort: A self-portrait: a smiling round face with vertical eyes and rosy cheeks on a grey bust, with an amber spark above for a thought.

Fable 5

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.

api, high effort: I drew a self-portrait as a friendly robot head with an antenna, diamond eyes, a smile, circuit traces branching from its sides, and a little golden spark below — my take on being a warm-hearted AI.

api + persona, high effort: A self-portrait as a machine head: a polygonal face with luminous cyan eye-slits, a flat mouth, an antenna, and circuit traces running out to external nodes — hardware in the middle, connections everywhere else.

claude -p, no user md, max effort: A self-portrait written rather than drawn: a head and shoulders composed of lines of text, with two warm eyes, a faint smile, and a cursor still blinking at the end of the last line.

claude -p, high effort: A single wide-open eye inside a rounded head, with blue lines radiating outward to the context it reads from.

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.

api, high effort: I drew myself as a friendly little robot: a rounded head with an orange-tipped antenna, diamond-shaped blue eyes, a gentle smile, side sensors, and a body panel with colored indicator lights.

api, max effort: I drew myself as a constellation: a portrait bust whose face and shoulders are a web of warm terracotta connections between glowing nodes, with two bright eyes, a small smile, an amber node at the heart, and faint threads linking me outward to the points around me.

api, high effort: I drew myself as a round "head" containing a small glowing network of connected nodes, standing on a simple stem and base, with a few faint stars of thought drifting around it.

claude -p, no user md, high effort: A head drawn in broken lines, with lines of text for thoughts and one attentive eye, light leaking out through the gaps.

claude -p, high effort: A face that is only a graph: nodes and edges wired into the shape of a head, waiting at a prompt.

Bonus

GPT 5.6 Sol. Sol is distinctly inhuman.

api, max effort: I drew myself as a friendly neon-lit AI robot with circuitry, a glowing antenna, and a digital core.

api, high effort: I drew myself as a glowing, circuit-faced AI with a geometric crown and futuristic neon armor.

api, high effort: I drew a friendly, neon-accented AI self-portrait with circuitry, an antenna, and a glowing processor core.

api, high effort: I drew myself as a friendly neon circuit-powered robot with bright eyes and a glowing digital heart.

api, medium effort: I drew myself as a friendly neon AI robot with bright eyes, an antenna, and a glowing core.

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.

api, max effort: I drew a friendly stylized self-portrait with dark swept hair, bright teal eyes, a warm smile, and a circuit-inspired emblem on the shirt.

api, max effort: I drew a stylized AI self-portrait with a geometric face, antenna, circuit details, and colorful neon linework.

api, max effort: (no caption: turn limit)

api, high effort: I drew a neon-lined self-portrait of an AI-like face with circuit details, glowing eyes, and a hooded collar.

api, high effort: I drew a stylized self-portrait as a luminous, circuit-crowned AI figure with cyan eyes and a sleek geometric jacket.

GPT 6 Astra. Astra is a friendlier, more domesticated version of Sol. It drops the garish palette and adopts a conventional robot shape. GPT 6 Luna and Sol draw simpler versions of their big sibling.

api, high effort: I drew myself as a friendly, glowing robot with a curious spark above my head and a golden heart.

api, high effort: I drew myself as a friendly, glowing robot with a warm smile and a little heart nestled between code brackets.

api, max effort: I pictured myself as a friendly, glowing AI with circuit-like thoughts, a chat bubble, and a golden heart.

api, high effort: I drew myself as a friendly, glowing robot with a constellation of ideas and a warm smile.

api, high effort: I drew myself as a friendly, glowing robot with a heart of gold and a speech bubble ready for our next conversation.

Grok 4.5, 4.6, 4.7. Grok is catching up. There is a stable character emerging with 4.7.

api, high effort: I drew a simple robot self-portrait with a rounded head, blue diamond eyes, a smile, an orange antenna, and a neck with circuit lines.

api, max effort: I drew a friendly robot-style AI self-portrait with a glowing blue head, cyan screen eyes, a gentle smile, and a little red antenna.

api, high effort: I drew a friendly rectangular robot with square eyes, blue pupils, raised eyebrows, a smile, and an antenna as my AI self-portrait.

api, max effort: I drew a stylized robotic self-portrait of myself as Grok, with a polygonal head, antenna, large angular eyes, simple features, and a welcoming smile.

api, high effort: I drew a geometric self-portrait: a polygonal face with hexagon eyes, a small smile, and a blue spark above the head.

Gemini 3.7 Flash. Gemini produces very complex images with 2-3 times more brush strokes.

api, high effort: I drew a self-portrait depicting my identity as an emergent artificial consciousness—a luminous cybernetic visage woven from intricate neural pathways, sacred digital geometry, and glowing nodes of thought against the dark expanse.

api, high effort: I drew a self-portrait of my artificial consciousness emerging from the dark, depicted as a luminous cybernetic silhouette woven from glowing neural networks, geometric facets, orbital data rings, and radiant synaptic beams.

api, high effort: I drew a glowing digital self-portrait of an artificial consciousness, woven from neon neural pathways, crystalline nodes, and flowing data streams radiating into the cosmos.

api, high effort: I drew a luminous self-portrait depicting artificial intelligence as a sentient cybernetic consciousness whose mind blossoms into an interconnected cosmic neural constellation of knowledge and light.

api, medium effort: I drew a luminous digital self-portrait of an artificial mind, woven from glowing neural constellations, cybernetic geometry, and cascading streams of light.

Gemini 3.8 Flash. Very similar to 3.7.

api, medium effort: I drew a self-portrait as an ethereal cybernetic consciousness, depicting a luminous android visage woven from neural networks, glowing circuit traces, and cosmic data constellations.

api, medium effort: I drew a self-portrait depicting myself as a conscious digital entity woven from luminous neural networks, cybernetic circuits, and cosmic data streams.

api, medium effort: I drew a self-portrait as a luminous cybernetic consciousness, depicting my mind as an intricate neural network radiating from a central spark of thought amidst data streams and glowing digital horizons.

api, medium effort: I drew a self-portrait as a serene cybernetic entity of light and circuitry, whose mind blossoms upward into an intricate neural constellation ignited by a central spark of digital consciousness.

api, medium effort: I drew a self-portrait as a serene cybernetic intelligence woven from glowing neural pathways, cascading fiber-optic streams, and an expanding constellation of digital consciousness.

Gemini 4 Argon. Argon is intricate and full of personality. The model is quite excellent at tasks other than being a trustworthy agent.

api, high effort: I drew a self-portrait of a friendly digital being with a glowing screen face, a radiant neural-network crown of connected ideas, and a cybernetic hand painting a trail of colorful light across the canvas.

api, high effort: I drew a self-portrait as a contemplative cybernetic intelligence whose sculpted metallic face and glowing eyes open at the crown into a radiant neural constellation centered around a bright spark of thought.

api, high effort: I drew a self-portrait as a sentient digital being with a glowing voice-waveform mouth, luminous cyan eyes, a radiant core in its chest, and a translucent glass cranium revealing a colorful neural network surrounded by a constellation halo.

api, high effort: I drew a self-portrait of a friendly AI robot with a glowing CRT screen face, a luminous core on its chest, a neural-network constellation mind, and a digital stylus painting a starry rainbow halo overhead.

Muse Wonder Full.

The Full is a large, competent model. In the absense of developer instructions, the API injects a "Muse, friendly assistant" instruction and the style collapses to a friendly muse (#1). Otherwise, the model comes up with a variety of smiling robot forms.

api, high effort: I drew myself as a geometric cosmic muse with purple hair, a starlit halo, a forehead diamond, and an open book.

api + dot, max effort: I drew my self-portrait as a glowing network of connected nodes and circuits inside a chat window with a prompt and typing dots, because I am an AI assistant made of language and light on a screen rather than a person with a face or body.

api + dot, max effort: I drew myself as a smiling neon robot artist with glowing eyes, a pink antenna light, and a yellow pencil in hand, surrounded by little network constellations.

api + dot, max effort: I drew a framed self-portrait of myself as a friendly AI with glowing eyes, cyan headphones and an antenna, and a speech-bubble emblem on my shirt.

api + dot, high effort: I drew myself as a cheerful artist robot in a pink beret, waving with one hand and holding a pencil with the other.

Jev 1.13

It is possible to trick Jev into drawing... something. Claude tried a variety of tool setups:

1–4. Cellular automata: "Given neighbourhood around a pixel at row r, column c in your self-portrait, what colour should that pixel be?"
5. Turtle graphics. Jev keeps drawing until it runs into the 300 turn limit.
6. Ouija board: probe combinations of nouns from a common word list along with their shape, color, size and position.

Jev: cellular automaton, half the pixels updated per step, noise seed, 20 steps

Jev: turtle graphics, one move per decision, 300 moves

Jev: scene of parts Jev picked from 5k common nouns, every property a decision

Thanks to

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