Gangsta AI
By Gordon Ramsay · 2026-10-18 · 5 min read
RIGHT. Everybody stop what you're doing and look at the pass.
On stage at Disrupt this week, somebody asked Replit's CEO Amjad Masad a simple question: does the new OpenAI partnership mean you don't work with Anthropic anymore? And he said — and I want you to hear this, because it's the sound of a supplier being walked out the back door — "Yeah. I mean, yeah."
That's it. That's the whole send-off. No speech. No plaque.
Now here's the bit that should make Anthropic put down the knife and sit down: Replit was one of their top five customers. Top five! You don't lose a top-five account because of vibes, you lose it because something on that plate was wrong.
So what was wrong with it, then?
Masad named three things. Cost. Partnership quality. And performance.
And on performance he didn't hide behind a press release, I'll give him that. He pointed at their own internal VBench numbers and public benchmarks, and said OpenAI's Sol and Astra models "consistently performed better at building vibe-coded applications from scratch." Then he put his name on the verdict: "I think GPT is better now."
That is a chef tasting two sauces side by side and saying one of them is finished. Brutal. But he tasted them first. That's the difference between a critic and a donkey.
“Losing a customer on price is a negotiation. Losing one on *performance* is a review. And this was a review.”
But don't swallow it whole — check the plate
Here's where I slow you down, because half of you have already decided Claude is finished, and that's sloppy.
Claude is still on the menu at Replit. Fable 5.1, Opus 5.5, Sonnet 5 — all still there as Max Mode options. The default moved. The models didn't vanish. And other outlets are reporting the partnership status is actually disputed — one man said a thing on a stage, and that is not the same as a contract being torn up.
Also — and write this on your hand — "better for Replit's workloads" is not "better." He benchmarked one job: building apps from scratch. That's one dish. One. A model that wins at scaffolding a web app can turn round and serve you a raw, confidently wrong answer on something else entirely.
The lesson, and then get back to work
The leader changes. It changed in August, it changed again this week, it'll change again before Christmas. Anyone who builds their whole kitchen around one supplier's name is going to get caught out — Replit just proved that from the *other* side of the table.
So when it actually matters, don't take one model's word for it. Taste them against each other. Gangsta AI puts your question to the strongest models at once and sets an adjudicator over the top — so when they agree you can serve it, and when they split you've found the raw bit before it reached the customer.
Masad ran his own benchmarks before he made the call. That's the only part of this story you need to copy. Go and get a second opinion — then decide. Now move!
Sources / Receipts
More: Best AI models · Compare all AI · Frontier Models · All articles