Minimal editorial cover: two diverging lines on a white background, one turning orange as it splits away

Replit Says It's Done With Anthropic, Bets Its Agent on OpenAI

CEO Amjad Masad told TechCrunch Disrupt that Replit no longer works with Anthropic, citing cost and performance — 'I think GPT is better now' — and revealed the coding platform was one of Anthropic's top five customers.

Product updatesReplitAnthropicOpenAI

Replit is no longer working with Anthropic. The company’s chief executive said so plainly onstage at TechCrunch Disrupt 2026 — and the reasons he gave were cost, partnership, and performance.

Asked whether the OpenAI partnership Replit announced in August meant the Anthropic relationship was over, CEO Amjad Masad answered: “Yeah. I mean, yeah.” The session recording was published by TechCrunch on October 16.

Masad said OpenAI’s models now perform better on Replit’s workloads. “I think GPT is better now,” he said, pointing to an internal benchmark called VBench and a public benchmark the company has published. OpenAI’s Sol and Astra models, he said, consistently performed better at building vibe-coded applications from scratch.

The split confirms the other side of a deal that has been forming since August. On August 18, Replit announced Free Mode and made OpenAI its primary model partner. Masad’s comments this week close the loop: the strategic relationship with Anthropic is over, even though Anthropic’s models — including Claude Fable 5.1, Opus 5.5, and Sonnet 5 — still appear as Max Mode options in Replit’s model-selector documentation.

The defection also came with a pointed aside about Anthropic’s business. Asked whether losing Replit had contributed to Anthropic’s reported revenue slowdown, Masad said Replit was one of Anthropic’s top five customers. The remark puts a rough number on the loss: a top-five API account is now directing its spend to a rival lab.

Replit’s own engineering data shows how far the OpenAI bet already goes. A September 29 engineering post documented GPT-6 Astra as the core loop of Replit Agent in Max mode, scoring 72% at $2.11 per task on the DeepSWE v1.1 benchmark and 49% at $2.53 per task on Terminal-Bench 4.0 — beating a sidekick architecture by 11 and 16 points. The company’s engineers wrote that Astra’s code quality let them use their code-review subagent less, with no drop in evaluation scores, describing a level of engineering quality they had not seen from any model before.

The timing matters beyond one vendor switch. AI coding platforms are among the largest buyers of model API capacity, and their partnerships are where lab rivalries turn into revenue. When a platform that processes millions of coding sessions moves its default workloads, the decision is made on benchmarks and unit economics — not loyalty. Masad framed it exactly that way: cost, partnership, and performance, with OpenAI the better partner “for now.”

Masad also used the session to argue that Replit is not about to be disrupted by the labs it buys models from. Asked whether OpenAI and Anthropic moving beyond code generation into building, deploying, and hosting apps threatens Replit, he said those offerings remain surface level next to Replit’s full-stack, integrated environment. OpenAI, he noted, relies on an external service — Cloudflare — for hosting, while Replit’s agents can spin up databases and key-value stores inside a production environment, and the company can deploy its entire stack on-premises for government and enterprise customers.

Whether “for now” holds is the open question. Model leadership in coding has changed hands several times in two years, and Replit has shown it will follow the numbers wherever they point. For this round, the numbers point to OpenAI.