Replit, the browser-based AI coding platform, is no longer working with Anthropic, according to co-founder and CEO Amjad Masad, who said so onstage at TechCrunch Disrupt on Thursday. The remark was notable because Replit had announced a partnership with OpenAI in August, but the company had given no public signal that its Anthropic relationship had ended.
Masad framed the change as a combination of factors rather than a single trigger. He cited cost, partnership considerations and performance, and said Replit relies on an internal benchmark alongside a public one to compare models. In his account, the models he referred to as Sol and Astra consistently did better at building vibe-coded applications from scratch, which he said made OpenAI the better partner "for now."
The practical substance of the relationship is that Replit sells a cloud development platform where users build and deploy apps inside a browser, and under the Anthropic partnership it ran that platform on Claude models. Masad said that arrangement had helped the company grow revenue significantly, which makes the switch a change to the infrastructure behind a product many freelancers and small teams use rather than a purely corporate announcement.
Pushback followed that claim rather than acceptance. Insiders acquainted with the Anthropic-Replit partnership objected to the comments, TechCrunch reported. The partnership remained in place, according to one individual who requested anonymity. What that individual described amounts to a source claim rather than a fact verified independently, and the CEO's onstage statement stands in direct conflict with it.
Evidence in the form of documents also points toward Anthropic usage in the recent past. Models identified as Claude Fable 5 and Claude Fable 5.1, both from Anthropic, were described in a September blog post as running within Replit Agent's production environment, the report said. What runs at present cannot be established from this. It does show, however, that at Replit, Anthropic models were in production as recently as September — a detail that fits awkwardly alongside a flat statement that the relationship is over.
TechCrunch said it reached back out to Replit for clarification and would update the story if the company responded. As of the report, that clarification had not arrived, leaving the central factual question — whether the partnership is terminated, winding down, or still active in some form — unresolved by the company itself.
Masad's onstage remarks went beyond the vendor question. He said enterprise is becoming a larger share of revenue: enterprise was about 2% of revenue last year, he said, and this year roughly 15% of revenue comes from enterprise contracts, with more than 50% of revenue coming from B2B overall. He did not share ARR figures, saying the company was not ready to do so.
He also described AI safety in terms of what agents need to operate. Comparing agents to humans who need a workable environment, he listed good orchestration, error recovery, safety and security, sandboxing, networking, and identity and authorization as the components of a habitat that would let AI be deployed safely. That is a stated position, not a tested result.
For freelancers, designers and developers who build on Replit, the immediate takeaway is that the model layer behind the platform may be in flux, and the company has not published a migration note, a deprecation timeline or a pricing change tied to the switch. Anyone with production workloads on Replit Agent has a reason to watch for official confirmation rather than treat a conference remark as a completed transition.
The cost angle deserves separate weight. Masad named cost first among his reasons, and model choice is one of the few levers a platform can pull to change its own margins without changing what it charges users. A platform that switches providers for cost and benchmark performance may pass savings through, absorb them, or change nothing visible at all; the evidence here does not say which.
The performance angle is similarly narrow. Masad's comparison rests on Replit's internal benchmark and a public benchmark, and on his characterization of two models as better at building applications from scratch. No methodology, sample, scoring rubric or result set was provided in the source material, so the claim should be read as a company's own evaluation rather than an independent finding about model quality.
The partnership angle is the least documented of the three. Masad said the decision reflected cost, partnership and performance together, but the evidence does not describe what changed in the commercial terms with Anthropic or what the OpenAI partnership announced in August actually covers. Readers should not infer that the August OpenAI announcement caused the Anthropic split; the source only notes the two events in sequence.
The dispute between Masad's statement and the unnamed insider matters for how this should be reported. A CEO speaking publicly is an attributable source; an anonymous person familiar with the arrangement is a weaker but still relevant one. When the two conflict on a basic fact, the honest position is that the status of the partnership is contested, not settled.
What remains unknown is substantial: whether Anthropic models are still serving any Replit traffic, whether the change affects all products or only Replit Agent, when any transition took effect, what it costs users, and whether the September production references have since been replaced. None of those details appear in the supplied evidence.
The useful conclusion for this audience is procedural. Treat the CEO's remark as a strong signal that Replit's model strategy is shifting toward OpenAI, treat the insider's denial as a reason not to assume the change is complete, and wait for Replit's own clarification before rewriting deployment plans or contracts around a provider switch that the company has not formally documented.