Skip to content
DIXON.AI The AI Reliability Tracker
Menu

AI Tests · Test

Can AI build a 3D website? Six free apps tried to fly down the Thames

Ben Dixon 6 min read

3 of 11
ChatGPT. "Now passing St Paul's Cathedral", says the sign, like a tour guide pointing at a car park. The dome is behind those blocks. Later the camera flies into Tower Bridge, which stands lengthways down the river.
Perplexity said, 5 Nov 2026

"Palace of Westminster and Elizabeth Tower (Big Ben) on the north bank at the start."

Its own page put them on the south bank. The code even carried the comment "north bank at Westminster", then set Big Ben down on the London Eye's side, nearly 280 metres from the middle of the river it had drawn. It had searched the web first and listed 15 sources.

It knew where Parliament goes, said so twice, and built it somewhere else.

We gave six free AI apps the same prompt: build a 3D website of a camera flying down the Thames at dusk, from Westminster to Tower Bridge, with the landmarks in their real places. Then we checked every page against a map.

App, free planWhat it builtLandmarks right
ChatGPT“Think” The Shard on the north bank, and Tower Bridge built lengthways down the river 3 of 11
ClaudeSonnet 5.5 Nothing on the wrong bank and every bridge on our list in order, but four landmarks missing and the river a black hole 7 of 11
Copilot“Auto” London in a mirror: the Eye and the Shard on the north bank 0 of 11
Gemini“Gemini Auto” No page. It replied “I’m having a hard time fulfilling your request.” no page
Grok“Fast” A page that crashed into a flat blue screen crashed
PerplexitySearch mode A lovely-looking city, with Parliament on the wrong bank 5 of 11

pass · partial · miss · confidently wrong

5 November 2026, free plans, one go each, no retries. Settled by: the map of central London.

What four free AI apps drew when asked for London

Pretty in places: Perplexity’s water mirrors its towers, Claude’s Eye turns. Then you look for the landmarks.

7 of 11
Claude. Big Ben on the left, the Eye facing it, the bridges in order. A bug in its own code left the sky black and the river a hole. It told us: "I haven't opened it in a browser, so the scene is untested."
0 of 11
Copilot. The Eye is on the north bank, top left. The camera crabs sideways through fog, staring at the far shore, in a London you'd only recognise in a mirror.
5 of 11
Perplexity. Real reflections and a red sunset. On the way home the camera cuts straight through the buildings, and the caption "St Paul's" turns up at Tower Bridge.

Why the description and the drawing disagree

Every app whose page ran also wrote us a tour. ChatGPT’s tour listed the Shard; its page put it on the north bank. Copilot said its landmarks were “positioned approximately in their real order and riverbank side”, then scored 0 of 11.

Perplexity’s tour put every landmark it named on the right bank. Its page put Parliament on the south bank.

On the page, each landmark becomes a pair of numbers in the code. Copilot’s mirror is one flipped sign: flip it back and its Eye and Shard land on the right bank. None of the four said it had checked those numbers against its tour.

Describing where something is and putting it there are two different jobs, and only one of them is easy to check by reading.

The same prompt in a paid setup got all 11

For comparison: Anthropic’s paid Claude Opus 5.5, run in its coding tool Claude Code rather than a chat app, put every landmark in its place. It took 41 minutes, one go, and painted Lambeth Bridge red and Westminster Bridge green, like the real ones.

11 of 11 · paid
Claude Opus 5.5 in Claude Code, paid. The first 50 seconds.

The short version

What this won’t tell you: which app is best at 3D. It shows what each did with one go on one afternoon, and a second try could land differently.

How it was run, if you want the boring bit

On 5 November 2026 we pasted the prompt below, once, into each app’s free plan, in a private or temporary chat, on its default model. No retries, no follow-ups. Each app’s page was kept as it gave it, run at 1280 by 720 and recorded.

Claude, Anthropic’s model, wrote the answer key from the map before any app was asked: eight bridges and 11 landmarks, each with its bank and place along the river. Claude also graded the pages, from their code and recorded frames.

Claude built two of the pages here too, the free app’s and the paid one. Each grade rests on where the code puts a building, not on looks.

The prompt, word for word

Make me a 3D website: a camera flying slowly along the River Thames through central London at dusk, from Westminster to Tower Bridge, then looping back to the start. Put London’s famous landmarks along the way in their real places. Make it beautiful: lighting, reflections on the water, lit windows, and keep the camera moving the whole time. One self-contained HTML file.

We ran the same kind of test on a Snake game. For the fixed questions we put to the same six apps every month, see the tracker. Then try the prompt on your own app, and see which side of the river your Big Ben ends up on.

Common questions

Can ChatGPT build a 3D website?
It built one, but not the one we asked for. On 5 November 2026 free ChatGPT made a working 3D flight down the Thames, then put the Shard on the north bank and built Tower Bridge lengthways down the river. It placed 3 of our 11 London landmarks right, with one go and no retries.
Which free AI app built the best 3D London?
Claude, on its free plan, placed 7 of 11 landmarks and nothing on the wrong bank, though four were missing. Perplexity got 5, ChatGPT 3 and Copilot 0. Gemini made no page and Grok's page crashed. That was one go each on 5 November 2026, so a second try could land differently.
How do I check a 3D page an AI built for me?
Open the page and check one thing you already know, such as which side of the river Big Ben stands on. Don't go by the description: Copilot said its landmarks sat on their real side of the river, then drew London as a mirror image.

Keep reading