Claude vs ChatGPT is the most searched comparison in AI, and almost every answer to it is a feature list. This page is not. Claude's two current models and ChatGPT's two current models were handed the same eight briefs, word for word, one attempt each, nothing fixed afterwards. What they built is playing below. Pick a brief, watch all four at once, open any build, and decide with your own eyes.
What is actually being compared?
Four models, not two chat windows. On the Claude side: Claude Opus 5.5 and Claude Fable 5.1, both at maximum effort inside Claude Code. On the ChatGPT side: GPT-6 Astra and GPT-6.1 Sol, both at extra-high reasoning inside OpenAI's Codex CLI. Each brief fixes the content, so the differences you see come from execution. The exact commands, the 150-minute cap and the recording scripts are on the methodology page.
Where do the differences show first?
In time and volume, before you see a pixel. On the landing page, Opus 5.5 worked for 43 minutes and Fable 5.1 for 28; GPT-6 Astra took 10 and GPT-6.1 Sol 9. That pattern holds on every brief: across the eight, the Claude models took roughly 2 to 11 times as long as the GPT models and wrote roughly 2.5 to 12 times as many output tokens. Whether that time buys a better build is what the recordings let you check.
The second difference is harder to see from a feature list. The two GPT builds keep landing on the same idea. Their landing pages are headlined "Less busy. More in flow." and "Less busy. More you.", share a terracotta accent and end the hero with the same strip of integration logos. Both physics simulations are called Kinetic. Both shooters are called Dead Signal and Deadlink, and both put a door marked Relay in front of you. Asked to invent any product at all, both GPT models invented a small gadget for using your phone less. The two Claude models converge too, on the dashboard and the kanban board, but far less often, which is useful if you plan to ask for two options.
Which builds the better website?
My read: the Claude pages look like a brand and the GPT pages look like a good template, and you will have to decide which of those you want. Opus 5.5 built a full booking widget into its hero, calendar and time slots included, and Fable 5.1 chose a single serif headline with the product shot underneath. Astra and Sol are cleaner and faster, and they are also more interchangeable. The open brief, where nothing was specified, is the sharpest test of taste.
Which builds the better dashboard or app?
On the analytics dashboard all four models computed identical figures from the fixed dataset, so the data handling is a wash and the craft is the whole story: Opus annotates the revenue peak and the average, Fable labels the best day, both let you flip the chart into a table, and the GPT builds choose calmer, greener palettes. The phone finance app and the kanban board test interaction, and all four models pass: every app sends $42.50 and updates the balance to $4,783.69, and every board drags, undoes and filters correctly. Claude Opus 5.5's app adds haptic feedback; the GPT boards add a friendly sidebar. The difference is time: Claude took 35 to 101 minutes on those two briefs, GPT 9 to 15.
Which builds the better game?
The shooter is where effort stops being an advantage. Opus 5.5 was still working when the 150-minute clock ran out, so its files were published as they stood; its arena is the most atmospheric of the four, a neon city at night, and its recording ends in a game over at 25 seconds. Fable finished in 75 minutes, Astra in 20 and Sol in 13. The GPT arenas are blockier and flatter, Sol's player also dies, at 35 seconds, and Astra survives the full 40-second script. The MOBA repeats it: two dense Claude lanes, both named Riftline, against two lighter GPT lanes.
How should you decide?
Not by reading this page: the honest answer to Claude vs ChatGPT is a comparison you run yourself, on your kind of work. Prices are public: Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, Fable 5.1 costs $10 and $50 (Anthropic's model list), while GPT-6 Astra is $10 and $50 and GPT-6.1 Sol is $2 and $10 (OpenAI's Sol announcement). The cheapest model here built a working shooter in 13 minutes; watch it before deciding that is not enough.
Four dashboards, labels hidden. Pick the one you would ship, then reveal who built it.
Take the blind test
Frequently asked questions
Is this Claude vs ChatGPT, or Claude vs GPT-6?
Both. ChatGPT is the app; the models inside it in October 2026 are the GPT-6 family. We ran GPT-6 Astra and GPT-6.1 Sol through OpenAI's own coding agent, Codex, and Claude Opus 5.5 and Claude Fable 5.1 through Claude Code, so each vendor's models work the way their developers actually use them.
Did anyone fix or polish the builds before publishing?
No. Each model got one attempt per brief and whatever it left on disk is what you see, bugs included. The only retries we allow are for infrastructure failures such as a rate limit, and those are listed on the build page with their reason.
Why don't you just say which one is better?
Because "better" depends on what you are building and what you value, and a score hides that. The facts we publish are measurements such as time, size and runtime errors. The judgement is yours, and the blind test exists so your judgement is not steered by a logo.
Can I try the builds, not just watch them?
Yes. Every finished build opens in a sandbox on its own origin, exactly as the model wrote it. The Open links under each video take you there.