Replit Ends Partnership with Anthropic, Cites Cost and Performance Reasons

Replit’s CEO Amjad Masad told TechCrunch Disrupt that the cloud‑based coding platform is no longer using Anthropic’s Claude models. He said the change followed an August partnership announcement with OpenAI and was driven by cost, partnership terms, and performance. According to Masad, internal benchmarks called VBench and publicly released tests show OpenAI’s Sol and Astra models outperform Claude for building new applications. He added that Anthropic was still a top‑five customer, but the company chose a different partner for now.

Despite Masad’s statement, insiders familiar with Replit’s recent communications say the Anthropic partnership remains active. Blog posts from late September list Claude Fable 5.1 and Opus 5.5 in Replit Agent’s production environment, and an employee X post praised Opus 5.5’s capabilities. A September 29 engineering post showed Fable 5 and GPT‑6 Astra running side‑by‑side, with Astra scoring 72 % on DeepSWE v1.1. These examples indicate Anthropic models were still deployed weeks after the announcement.

Replit’s shift comes as the company’s enterprise contracts grow from about 2 % of revenue last year to roughly 15 % this year, with more than half of revenue now B2B. Masad highlighted cost‑optimized “Power Mode” and high‑performance “Max Mode” plans that use different model families. He also emphasized the need for an AI “habitat” with orchestration, sandboxing, and security scans before enterprise deployment, a framework Replit has been testing since launching Agent 1 in 2024.