Amjad Masad, CEO and co-founder of Replit, said that the company no longer works with Anthropic in remarks he made while appearing onstage at the TechCrunch Disrupt event on Thursday. However, sources familiar with the partnership provided a different account, while materials published on Replit’s website, blog, and social media accounts indicate that Anthropic models continued to be used until only a few weeks before the statement.
A Decision Linked to Cost and Performance
Masad explained that the decision was made for several reasons, including cost, performance, and the nature of the partnership. He said that Replit uses an internal benchmark to evaluate performance indicators, and also publishes a public benchmark, and that the Sol and Astra models had, according to the company’s evaluation, consistently performed better at building applications based on prompt-driven coding from scratch. He added that ChatGPT is currently better, and that these factors led Replit to consider OpenAI the better partner for the time being.
Replit announced a partnership with OpenAI in August, but the announcement did not include anything indicating that its cooperation with Anthropic had ended. Replit’s cloud platform had been using Claude models to power its services, and the partnership was said to have helped the company significantly increase its revenue.
Conflicting Evidence About the Partnership’s Status
A person familiar with the partnership between Anthropic and Replit, who requested anonymity, rejected Masad’s assertion that the cooperation had ended and said that the partnership was still active. A blog post published by Replit in late September also referred to the Claude Fable 5 and Claude Fable 5.1 models as operating in Replit Agent’s production environment. During the same period, a Replit product employee posted on X about using the Opus 5.5 model and promoting its capabilities.
TechCrunch requested additional clarification from Replit and said it would update the story if it received a response. Therefore, the source does not establish that Replit’s transition from Anthropic to OpenAI has been completed in full; rather, it establishes the existence of conflicting statements and chronological evidence of continued use of Anthropic models.
What Does This Mean for AI Coding Platforms?
The incident reveals that choosing an AI model within development platforms depends not only on technical capability; cost, partnership terms, and internal test results are also influential factors. The continued appearance of multiple models in a single production environment may also mean that the transition between providers is gradual or that the platform uses more than one provider, but the source does not provide enough detail to settle the matter.
Masad also discussed the growth of Replit’s business, saying that enterprise contracts accounted for about 2% of revenue in the previous year, while they reached about 15% this year, and that more than 50% of revenue comes from the business sector. He did not disclose annual recurring revenue figures.
Masad stressed that deploying AI agents safely requires what he described as an integrated operating environment encompassing orchestration, error recovery, safety and security, isolation within sandbox environments, networking, and identity and authorization management. This point remains separate from the status of the partnership, but it clarifies the operational requirements that Replit considers necessary for using agents in software development.