Replit CEO says Anthropic relationship has ended, but that claim is disputed

Replit CEO Amjad Masad says the AI coding company has stopped working with Anthropic, citing cost, partnership and performance. But the split is disputed: a person familiar with the relationship told TechCrunch it remained in place. Replit’s own engineering records also document Claude running in production as recently as September.

The disagreement concerns a supplier behind Replit’s core product. Replit lets users build and deploy applications through their browsers, and it used Anthropic’s Claude models under the companies’ partnership. Masad made his remarks onstage at TechCrunch Disrupt, according to TechCrunch’s October 16 report.

Masad’s case for OpenAI

Replit announced an OpenAI partnership in August. Masad now describes that company as the better partner for Replit, rather than simply another model supplier. “I think ChatGPT is better now,” he said, explaining that the choice reflected both commercial considerations and his assessment of performance.

Masad said Replit uses an internal benchmark to assess performance and also publishes a public benchmark. He said Sol and Astra consistently perform better at building applications from scratch using natural-language instructions—the practice often called vibe coding.

His explanation was also explicitly provisional: OpenAI was the better partner “for now.” The rationale combined cost, the partnership itself and model performance. It was not a claim that the benchmark result alone determined Replit’s choice.

A dated record of Claude use

Replit’s September 29 engineering post, Free the models: Harness design at the frontier, supplies a concrete historical record. Its production comparison covered one week per model: Claude Fable 5 in August, followed by Fable 5.1 and OpenAI’s GPT-6 Astra in September. These were described as production observations, not merely laboratory trials.

That establishes recent Claude use, not continued use on October 16.

Different models, different jobs

The same engineering post explains why Replit evaluated multiple models. Its stated approach was to let models work differently, with quality at minimum cost as the goal. Replit argued that the strongest coding model need not also be strongest at designing interfaces, making slides or writing emails.