Amjad Masad had been onstage at TechCrunch Disrupt for most of an hour on Friday when he was asked about the models inside his company’s product. Replit, the coding platform he co-founded, is no longer working with Anthropic, he said. The reasons were cost, the partnership and performance. “I think ChatGPT is better now.”
The line landed harder than the rest of the interview. Replit had built its coding agent on Anthropic’s Claude models, and the relationship had been credited with a year of fast revenue growth at a company that sells software development to people who do not write software. Within hours, people familiar with the matter were pushing back, saying the partnership remained in place. Neither company resolved the contradiction when asked.
What Masad described is a supplier change, ordinary in an industry where model choice has become a procurement decision. What made it notable was the setting and the dispute attached to it. A chief executive announced a breakup from a stage, and the other side’s camp said there had been none.
The technical history is not in dispute. Replit runs a browser-based development environment where a user describes an application and an agent writes, tests and deploys it. Anthropic’s models did much of that work.
A Replit blog post in late September listed Claude Fable 5 and Claude Fable 5.1 among the models running in the production environment for Replit Agent, and a Replit product employee wrote on X about using Opus 5.5. Replit also announced a partnership with OpenAI in August, without saying at the time that anything had changed with Anthropic.
Masad’s account of the decision was specific about method. Replit runs an internal benchmark, he said, and publishes one as well, and two models he identified as Sol and Astra came out ahead at building applications from scratch. He described the call as a judgment across cost, partnership and performance, and said the other side had been a better partner “for now.”
The phrase matters as much as the claim. Replit is not a lab and does not publish its evaluation harness in a form outside researchers can reproduce, so the benchmark comparison is a statement of internal preference rather than a result anyone else can check. A platform that routes work across several model suppliers can also change its mix without changing what the customer sees, which makes a public announcement a marketing act as much as an engineering one.
The business mix gives the timing some context. Masad said enterprise contracts accounted for about 15 percent of revenue this year, against roughly 2 percent last year, and that more than half of revenue now comes from business customers. He did not give an annual revenue figure. Enterprise buyers ask different questions than individual developers: which model runs the code, where the data sits, what happens when a supplier raises prices or withdraws a model.
His description of what an agent needs was more revealing than the vendor comparison. Agents require a habitat, he said, meaning orchestration, error recovery, safety controls and sandboxing. That is a statement about where the engineering work sits. The model is one input; the runtime around it decides whether a generated application deploys, breaks, or quietly bills a customer for a database it did not request.
Cost is the pressure behind most of these decisions. Inference is the cost of goods sold for a coding agent, and an agent that iterates on its own output burns tokens in volumes that a chat product never approaches. A few points of gross margin decide the pricing a platform can offer, and model suppliers have been competing on exactly that margin for two years.
The cost of switching has fallen at the same time. Teams that once wrote directly against one supplier’s interface now put a routing layer between their product and the models, keep an evaluation set, and send each task to whichever supplier scores best on it. The abstraction is imperfect, because models differ in how they follow instructions and in what they refuse, but it is good enough that a platform can move a workload in weeks rather than quarters.
That has changed how contracts look. Model agreements now carry the features of cloud commitments: volume tiers, notice periods, credits, joint marketing. The word partnership covers several arrangements, from a reseller agreement to a company that happens to be a large customer, and the difference is rarely visible from outside. A partnership that ends may be a contract that expired, a routing change, or a relationship that was never exclusive.
For buyers, the ambiguity is the problem. A customer signing a multiyear contract wants to know which supplier sits behind the product and what notice it gets if that changes. Two companies describing the same relationship differently, and declining to say more, leaves that question open at the moment a sales team is trying to close deals.
Analysts who follow developer tools said the reliable signal will be Replit’s own documentation, model list and pricing page over the next several weeks, and whether Anthropic’s name reappears in release notes. Announcements are cheap; production environments leave records.
Anthropic spent the same week on matters of a different scale, pricing a large chip financing and signing its first Canadian data-center lease. Those commitments rest on assumptions about demand from exactly the kind of customer Replit represents: companies that resell model capability inside a product. One departure is a data point. A pattern of them would be something else.
What the episode does not settle is whether Replit has left, reduced its use, or simply added a second supplier and described the change in the plainest terms available to a founder on a stage. The next version of the product will answer it, and so will the next earnings season for whoever is counting.
