# Replit CEO Warns AI Application Companies Against Dependence on a Single Model Provider

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/23502
Published: 2026-09-24T01:59:11.000Z
Updated: 2026-09-24T01:59:11.000Z

Replit CEO Amjad Masad urged AI application companies to preserve flexibility as model pricing, performance and capabilities shift across a rapidly changing market.

“The most important thing is optionality,” Masad said.

Replit has shifted its model usage among several providers, including a period when Google received more token volume than Anthropic. Anthropic served as Replit’s “workhorse” for more than a year, while Google’s Gemini models delivered strong price performance for some tasks.

The company now uses models from every provider and has sometimes built models internally, Masad said. He added that the balance between developing proprietary models and relying on external providers can change every three to six months.

That approach addresses a concentration risk for AI application companies. A model provider can change its prices, improve its own products or make a capability available directly to customers. Replit aims to compete through its infrastructure, proprietary benchmarks, evaluation methods and expertise in routing models to specific tasks.

Replit is positioning itself as an agentic software-creation platform rather than only an interface for a single AI model. Its products cover application creation, code execution, deployment and database management.

Replit Agent launched in September 2024 as a system that could create and deploy applications. On July 21, 2025, Replit introduced separate development and production databases to make testing and deployment safer. Agent 3 can test and fix code while creating custom agents and workflows.

Replit raised $250 million at a $3 billion valuation on Sept. 10, 2025, in a round led by Prysm Capital. The company’s annualized revenue rose from $2.8 million to $150 million in less than a year.

Masad said Replit’s product can outperform a model provider’s own product in some use cases because Replit evaluates and integrates models around specific user problems.

“But if you focus on cost at the expense of performance, you’re going to lose,” Masad said.