GPT 5.6 vs Claude Fable 5: we stress tested both models in one afternoon

It was a Saturday, the office was empty, and GPT 5.6 had just landed. Can and I had been waiting for a quiet moment like this. 2 models on 2 screens, both set to the same medium intelligence tier, and 3 briefs we had prepared to make them sweat. We said we won't edit, we won't make corrections and only 1 shot each. We agreed to score every test out of 10, separately, and only compare numbers at the end.

After 24 years in this industry, I have learned one thing about tools: the claims mean nothing until you taste the food yourself. So we did.

How we tested GPT 5.6 against Claude Fable 5

Brief 1 was the torture test: a pricing page with 3 tiers, a monthly and annual toggle that recalculates prices live, and a comparison table underneath. We counted down, hit enter on both at once, and watched. Claude finished first, by under 10 seconds. Close enough that speed told us nothing yet. What caught my eye was the shape of the answers. Claude replied inline, right there in the chat. GPT produced a complete HTML file with branding we never asked for. I made a note of that and moved on. It would come back later.

Can AI clone the Apple Weather app?

Brief 2 was the one I was nervous about. A single screenshot of Apple's Weather app from my MacBook, and 7 words: recreate this screen pixel for pixel, fully interactive. GPT came back in about 2 minutes, and honestly, it was accurate. 10 years ago this work happened bit by bit and could take a week. Claude took 5 minutes, and while some padding and margins drifted, it added a small rain animation nobody requested, and the interactions were the best in the room. Icons were rough on both. That gap is still ours to fill, and it takes a designer minutes.

The agentic test that decided the winner

Brief 3 was simple on purpose: find 3 record shops near Shoreditch, pull their hours and specialties, compare them, no corrections allowed. This is where the note from brief 1 came back. Claude answered like a person would: hours, ratings, and a working map, straight in the chat. GPT did what it had done all afternoon and handed us another HTML file. Sitting there, I asked the question out loud: I am asking you a question, why are you giving me a website? GPT behaves as if every prompt is client work waiting to be built. Sometimes that is exactly what you want. On a Saturday afternoon asking about record shops, it is not.

Which AI model should designers use in 2026?

We compared scores. Can had Fable 5 at 23. So did I, without knowing his numbers. GPT took 21 and 22. A 2 point gap after 3 tests is close, and the whole difference came from that third brief. Accuracy plus context beat raw speed.

My honest advice: do not take our verdict, or anyone's. Pick 3 briefs from your own work, run them through both models, and score them out of 10. It costs you 20 minutes and it will tell you more than any comparison video, including ours.