A top-five Anthropic customer just walked out. The Replit Anthropic split went public at TechCrunch Disrupt, where Replit CEO Amjad Masad told the audience his company is “no longer working with Anthropic” — and that ChatGPT now powers Replit Agent by default.
Stunning reversal: Replit used to run on Claude, and Masad once said the company was one of Anthropic’s five biggest customers. Two months after the August 2026 OpenAI partnership, Replit crowned ChatGPT the winner of its internal bake-off — on cost, partnership, and performance.
What Replit’s CEO actually said onstage
Masad did not hedge. The money quote: “I think ChatGPT is better now.”
That verdict came from two benchmarks — Replit’s own VBench plus a public benchmark — which found that OpenAI’s Sol and Astra models “perform better on building vibe-coded applications from scratch.” The models that ship working apps fastest won the contract; brand loyalty lost. Read the full onstage account in TechCrunch’s report on the Replit CEO’s announcement.
Why benchmarks beat brand loyalty: the evidence
How a top Anthropic customer became a ChatGPT shop:
| When | What happened | Why it matters |
|---|---|---|
| August 2026 | Replit announces a partnership with OpenAI | While Claude still ran production |
| Late September 2026 | Replit blog cites Claude Fable 5 and 5.1 in Replit Agent production | Anthropic models were live weeks before the claim |
| October 17, 2026 | CEO says “no longer working with Anthropic” at TechCrunch Disrupt | The public breakup |
| Days after | Masad clarifies Anthropic models remain accessible; ChatGPT is the default | A “default flip,” not a hard rip-out |
| Same weekend | An insider tells TechCrunch the partnership is still in place | The story is disputed |
Three numbers make this real:
- A top-five customer walked. Volume does not guarantee loyalty when benchmarks flip.
- Two benchmarks agreed. VBench plus a public benchmark both ranked OpenAI’s Sol and Astra higher for vibe-coded apps built from scratch.
- Enterprise revenue went from 2% to 15%. Shared on the same stage — the number that explains why Replit now cares about benchmark bragging rights.
The twist: did Replit really cut ties?
Not so fast. An insider told TechCrunch the Anthropic partnership is still in place — contradicting the CEO — and pointed to a late-September Replit blog post citing Claude Fable 5 and 5.1 in production, plus a Replit employee posting on X about Opus 5.5.
Masad’s weekend statement tried to thread the needle: “Replit remains model agnostic… OpenAI is now our preferred partner for powering Replit Agent, based on performance, economics, and the strength of our partnership.” See the coverage of the insider dispute over whether Replit really dropped Anthropic for the full back-and-forth.
The builder’s lesson: smart companies never burn the bridge. Replit flipped the default but kept Anthropic accessible. That is what “model agnostic” means in production.
The bigger money story: enterprise went from 2% to 15%
Buried in the same appearance was the number that matters for anyone selling AI: Replit’s enterprise contracts grew from 2% to 15% of revenue, with B2B now over 50% of its business.
That explains the timing. Consumers buy “it feels magical”; enterprises buy benchmark receipts. A buyer signs when you show VBench numbers and a public benchmark ranking your stack first — not when you name-drop a model brand. Vibe-coding is moving upmarket.
How to profit from this split: 5 concrete steps
- Build your own VBench. Pick 10 tasks your customers actually do, score every frontier model monthly, and switch when the winner changes.
- Price per task, not per token. Track completed tasks per dollar and let that number choose your provider.
- Stay model-agnostic. Put one interface between your app and the model so a default flip costs an afternoon, not a rewrite.
- Negotiate like a top-five customer. Partnerships are deals — price, support, co-marketing. Growing usage is leverage.
- Sell the proof. Publish your own benchmark results to win enterprise buyers — Replit’s 2%-to-15% shift shows where the money is.
Should YOU switch AI providers? The Replit test
1. Run your own evals
Score models on your exact use case, like Replit’s VBench — not public leaderboards.
2. Price per task
A cheap model needing 3 retries is the expensive one. Measure completed tasks per dollar.
3. Check the deal
Replit cited economics and partnership strength. Better terms can flip the winner.
4. Flip, don’t rip out
ChatGPT became the default; Anthropic stayed accessible. Never burn the bridge.
5. Sell the proof
Benchmarks are marketing gold. Enterprise buyers pay for proof, not vibes.
Frequently asked questions
Is Replit no longer working with Anthropic?
That is what CEO Amjad Masad said onstage at TechCrunch Disrupt, citing cost, partnership, and performance — but an insider disputed it, and Replit later clarified Anthropic models remain accessible.
Can I still use Claude models on Replit?
Yes — Masad clarified customers can still access Anthropic models; ChatGPT (OpenAI’s Sol and Astra) is simply the new default for Replit Agent.
Why did Replit switch to OpenAI?
Cost, partnership strength, and performance — backed by Replit’s internal VBench and a public benchmark favoring OpenAI’s models for vibe-coded apps built from scratch.
What is Replit’s VBench benchmark?
VBench is Replit’s benchmark for scoring AI models. Alongside a public benchmark, it drove the call to make ChatGPT the default engine of Replit Agent.
Is the split confirmed or disputed?
Disputed. The CEO announced it onstage, but an insider says the partnership continues and Replit’s late-September blog cited Claude Fable 5 and 5.1 in production. Replit’s line now: “model agnostic,” OpenAI the preferred partner.
What should AI builders take away?
Benchmark your own use case, price per completed task, and stay model-agnostic so you can flip providers like Replit did. And follow the money: Replit’s enterprise revenue grew from 2% to 15% — buyers pay for proven performance.
Keep Reading