Replit CEO Says Company Has Stopped Working With Anthropic, Citing Cost and Performance

News Summary

Replit, the browser-based software development platform, has stopped working with Anthropic, according to co-founder and CEO Amjad Masad. Speaking onstage at TechCrunch Disrupt on Thursday, October 16, 2026 (Pacific Time), Masad said the company made the change for a mix of cost, partnership and performance reasons. TechCrunch reported the remarks at 2:22 PM Pacific Time on October 16, 2026. The claim is notable because public signs of the relationship were still visible only weeks earlier, and at least one unnamed insider disputes it.

What the CEO Said

Masad told the audience that Replit no longer works with Anthropic. "It was a decision based on both cost, partnership, and performance," he said. On the quality of the models, he was direct: "I think ChatGPT is better now," and he added that "Sol and Astra perform better on building vibe-coded applications from scratch."

He said Replit evaluates models with an internal benchmark and also publishes a public one, which is how the company compares candidates for the app-building tasks at the center of its product. Vibe coding, the practice of describing software in plain language and letting an AI system write it, is the core use case Replit serves.

The Agent "Habitat" Argument

Beyond the vendor change, Masad spoke about what AI agents need to do useful work. "Agents need a habitat for them to exist and do good work," he said. In his view that habitat requires strong orchestration, error recovery, safety, security, sandboxing, networking, and identity and authorization. This is an engineering-education point worth noting: a capable model is only one part of a working agent system, and the surrounding infrastructure largely decides whether generated software can be run and deployed reliably.

Background on the OpenAI Relationship

Replit announced a partnership with OpenAI in August 2026. Separate coverage dated August 19, 2026 describes Replit launching a "Free Mode" powered by OpenAI's GPT-5.6 Luna model after an 80% price cut on that model. It was offered to subscribers on the roughly $20 per month Core tier and $100 per month Pro tier, and it can switch automatically to more capable models when a task needs them. Replit president Michele Catasta described the launch as the first of many joint efforts. We did not find published financial terms or exclusivity details for the deal.

Before Masad's comments, there had been no public indication that Replit had left Anthropic. Earlier reporting, including a January 2025 Semafor interview, noted that Replit's Agent product became practical after Anthropic's Claude 3.5 Sonnet posted a record score on the SWE-bench coding evaluation, which shows how central Anthropic's models were to the product's early growth.

Conflicting Evidence

TechCrunch noted that the picture is not fully clear. An unnamed insider said the Anthropic partnership was still in place. Replit's website, blog posts and social media showed use of Anthropic products as recently as a few weeks before the report. A late-September Replit blog post named Claude Fable 5 and Claude Fable 5.1 as models running in the production environment of Replit Agent, and around the same time a Replit product employee posted on X about using Anthropic's Opus 5.5 model.

TechCrunch asked Replit for clarification and said it would update its story if the company responded. Until then, it is unclear whether "no longer working with" means a complete end to all Anthropic usage, a shift of the main partnership, or a change in which models power default features. Readers should treat the details as developing.

Business Figures Shared

Replit did not disclose annual recurring revenue. Masad did say that enterprise accounted for about 2% of revenue last year and about 15% this year, and that more than half of revenue now comes from business customers (B2B). The shift suggests Replit is moving from individual learners and hobbyists toward organizations that build internal tools.

Why It Matters

Model choice is now a competitive decision that platforms revisit as quickly as new releases arrive. Replit's public benchmark and Masad's emphasis on cost and performance show how application builders weigh price per task alongside raw capability. For students and developers learning to build with AI, the lesson is practical: evaluate models on your own tasks, keep your architecture flexible enough to swap providers, and invest in the surrounding systems (sandboxing, security, error recovery) that make agents dependable.