Replit Cuts Ties With Anthropic, Standardizes on OpenAI for Agent
Replit is no longer working with Anthropic, co-founder and CEO Amjad Masad said onstage at TechCrunch Disrupt. TechCrunch
Masad cited cost, partnership and performance as the basis for the decision. Prior to its deal with OpenAI in August, Replit relied on Anthropic's Claude models to run its platform. The company had been, until recently, a top-five customer of Anthropic.
In production terms, the switch is complete. OpenAI's ChatGPT is now the default model on Replit's platform. OpenAI is now Replit's preferred partner for powering Replit Agent. That distinction matters. Customers can still access Anthropic models on Replit's platform, and the company says it remains model agnostic, with users able to choose Anthropic's Claude or other supported models.
The change follows a busy shipping period for Replit. On August 18, 2026, the company announced Free Mode to expand what is possible with AI, stating that with OpenAI, Replit helps users accomplish more every day. On September 10, 2026, Replit said its Databricks integration is generally available with native Lakebase support. On September 25, 2026, Replit announced it is acquiring Atta to add interactive charts in chat.
On September 29, 2026, Replit published work on agent harness design under the title Free the models: Harness design at the frontier, stating Replit Agent lets the model choose its subagents' tier and effort. On October 13, 2026, Replit introduced Replit Solutions Partners to help organizations put code to work. On October 16, 2026, Replit announced Bring Your Own Cloud, offering enterprise isolation and control.
The broader context here is how agentic coding platforms negotiate defaults. Model agnostic does not mean model neutral in practice. There is the model router and model picker users see, and there is the tuned path underneath, the harness, evals, tool use, retry logic and cost controls optimized around a preferred model for Agent loops. Changing that preferred path is operational work, not a config flag. It affects token burn, latency under multi-step tasks, and how subagent delegation behaves when the primary model decides tier and effort.
Looking at what this means for platform operators, inference economics dominate once agents run at scale. A top-five customer leaving a lab is leverage in both directions. For the application layer, lower unit cost and tighter co-engineering on the default path can fund more generous free tiers and longer agent runs. For the model layer, losing a high-volume coding workload removes a dense source of real-world tool-use traces. Coding agents generate structured, verifiable, multi-turn workloads that labs use to harden function calling and planning. That feedback loop is valuable.
In this author's view, the enterprise moves deserve equal attention. Bring Your Own Cloud separates the control plane from the model plane, giving customers isolation while Replit retains the agent experience on top. Pair that with a Databricks integration with native Lakebase support and a services partner program, and the pattern is familiar from previous platform shifts. Win developers with a fast default, then add deployment isolation and integration points to carry the same workflow into production data environments. Worth flagging for buyers is the split to test in procurement, default behavior versus selectable models, data residency under BYOC, and how evals track Agent quality as the underlying preferred model evolves.
What this enables, if the cost and performance trade holds, is more agent time per dollar. That is the constraint that has limited agentic development environments so far, not raw code generation quality. Cheaper, well-integrated runs allow longer planning horizons, more tool calls, and more iteration before human review.
