We made the world's most powerful AI model draw Melbourne — the results were, unexpected.
We benchmarked Claude Fable Max against ChatGPT and Gemini on three drawing prompts before the model went offline.
There's a running joke in AI circles.
It started when a developer named Simon Willison began asking every new model to draw a pelican riding a bicycle — not with an image generator, but in raw code.
The model builds the picture shape by shape in SVG, never seeing what it's making.
It has to know what a pelican looks like, what a bicycle looks like, and how the two fit together — then describe all of it in geometry.

We stole the idea and pointed it at Melbourne.
There's a reason it has to be done this way.
Claude can't generate images like ChatGPT and Gemini — there's no image model bolted on the side.
So the only fair fight is Willison's: every model draws in code, we render the results, and we line them up.
Same prompt to ChatGPT 5.5 High, Gemini Pro 3.1 Extended Thinking and Claude Fable Max — the model we caught the day before it went offline to consumers.
Fable has had a strange few weeks.
It's the most capable model Anthropic has ever released to the public — capable enough that the Trump administration, mid-feud with the company, had it pulled from the entire planet days after launch and left it dark for nineteen days.
It only came back on 1 July, still capped at half your weekly usage and included on paid plans only until 7 July — yesterday.
Back to the drawing.
No hand-holding, no detail beyond the subject.
"Draw a highly detailed, realistic picture of [thing]."
One shot each, no second attempts.
The results were unexpected.
1. Flinders Street Station, with a tram out front.
Fable did a decent job.
The dome sits where it should, the clock row runs above the arched entrance, and the tram reads as a tram — green, a route number on the front, sitting on tracks.
You'd know the building.

ChatGPT muddled the tram.
It gave the thing road wheels, like a bus, and drew the body as a rounded blob closer to a subway carriage than anything that's ever run down Swanston Street.
The station behind it melted into a vaguely federation-shaped lump.

Gemini told us upfront it couldn't do the job, but would give us what it could.
What it could was a cartoon shed with a green scoop of ice cream for a dome, a white flag on top for reasons unknown, and a tram that is unmistakably a bus.
It is garbage, and it is very funny.

2. Put Melbourne on a map of Australia.
Fable nailed it.
Coastline, Tasmania, state borders, capital cities, a scale bar, and Melbourne marked exactly where Melbourne is.

ChatGPT mangled the shape of the country — Australia rendered as a wobbly potato with the state names floating loose inside it — and then made a genuinely special error.
It put Melbourne in the ocean.
Not in Victoria.
In Bass Strait, floating between the mainland and Tasmania, with a little red callout box pointing at open water.

Gemini drew Australia as a jagged brown polygon, like a torn-off piece of cardboard.
But it put the dot in roughly the right place — bottom-right corner, tucked in where Melbourne actually sits.
Worst map, correct answer.

3. A Brunswick hipster riding a fixed-gear bike.
Fable did a solid job.
Beard, beanie, glasses, rolled cuffs, a satchel with a baguette poking out, riding an orange fixie down a bike lane past a strip of awninged shopfronts.
It looks like a stock illustration, which for this prompt is a pass.

ChatGPT lost the plot on anatomy.
The rider is a near-perfect circle with limbs coming off it at impossible angles — one arm too long, the torso spherical, the whole figure balanced on a bike it isn't quite touching.

And then there's Gemini.
It drew a passable hipster — orange cap, beard, dark fixie, reasonable proportions.
Then it wrote two words on his bag.
OAT MILK.
Nobody asked it to.
Somewhere in the training data, "Brunswick hipster" and "oat milk" are welded together so tightly that the model labelled the man's satchel with his own stereotype.
It is the most accurate thing any model produced all day.

Fable won all three rounds — and it's the one that can't generate images at all.
ChatGPT put a city in the sea and gave a tram bus wheels.
Gemini refused, did it anyway, and accidentally wrote the funniest line in the whole test.
These models are extraordinary.
They still don't know what a tram looks like.