Replit CEO says company stopped working with Anthropic

Replit’s relationship with Anthropic is disputed. On October 16, 2026, TechCrunch reported that CEO and co-founder Amjad Masad said Replit was no longer working with Anthropic; the same report cited an unnamed person familiar with the partnership who said it remained in place. The date is when TechCrunch published its account, not an established date for Masad’s remarks. TechCrunch reported both accounts.

What Masad said about Anthropic

Masad cited cost, partnership and model performance as reasons for Replit’s reported shift toward OpenAI. He said Replit’s internal and public evaluations found OpenAI’s Sol and Astra models performed better at building applications from scratch. That is Masad’s account of the evaluations and the company’s reasoning.

Claude models in Replit

Replit’s engineering post, published September 29, 2026, said Claude Fable 5 was in Replit Agent production in August, and Claude Fable 5.1 was in production in September. Replit’s model selector also lists Claude models in Power and Max modes, alongside OpenAI models and, in Max mode, Kimi K3. Replit’s engineering post and model-selector documentation give the production and selection details.

Core and Pro users can manually choose from models available to them in Power or Max mode. Availability varies with rollout, organization settings and authorization. Free Mode uses intelligent routing rather than manual model selection.

What Replit’s benchmarks measure

In its September 29 post, Replit reported that Replit Agent in Max mode, with GPT-6 Astra as its core loop, scored 72% on DeepSWE v1.1 at $2.11 per task and 49% on Terminal-Bench 4.0 at $2.53 per task. Each score was the mean of four repetitions. The DeepSWE run covered 113 tasks; the Terminal-Bench run covered 63, with three GPU tasks excluded.

Replit also reported results for GPT-6 Astra running in mini-swe-agent and for a sidekick setup using one long-lived worker. On DeepSWE v1.1, the low-effort mini-swe-agent run scored 67% at $1.60 per task, the xhigh-effort run scored 74% at $4.43, and the sidekick setup scored 61% at $1.34. On Terminal-Bench 4.0, the corresponding results were 42% at $2.25, 60% at $5.86 and 33% at $1.84.

Those published configurations are an Astra-based Replit Agent, Astra in mini-swe-agent at two effort settings, and the sidekick setup. Claude was not one of the configurations in these results, so the scores describe different agent setups rather than a direct Astra-versus-Claude test.

Replit’s earlier OpenAI integration

On August 18, 2026, Replit announced Free Mode, powered by GPT-5.6 Luna. Replit’s announcement describes that OpenAI integration; Masad’s remarks concern the separate, disputed relationship with Anthropic.