Software · 17 October 2026 · 4 min read
Replit cuts ties with Anthropic in favour of OpenAI: inside the coding agent showdown
In brief: Speaking onstage at TechCrunch Disrupt, Replit CEO Amjad Masad revealed that the development platform has ceased working with Anthropic, shifting its primary workload to OpenAI. Driven by lower token costs and superior benchmark results from models like Astra and Sol in vibe coding, the split illustrates how developer-tool platforms are rapidly rethinking model partnerships to sustain agentic software creation.
by Team Mocchi's
For more than a year, the partnership between Replit and Anthropic served as the showcase deployment for the Claude model family in software engineering. That collaboration, which helped accelerate the cloud platform's rapid revenue growth, has now come to an abrupt end. Speaking onstage at TechCrunch Disrupt in San Francisco, Replit co-founder and CEO Amjad Masad announced that the company is no longer working with Dario Amodei's AI lab, having redirected its primary workloads to OpenAI.
The strategic shift marks a critical development in the battle over vibe coding, the workflow where users describe features in plain natural language while autonomous agents assemble, debug, and deploy the entire application stack.
Performance and economics: the internal benchmark verdict
Masad framed the departure not as an ideological break, but as an empirical outcome dictated by cost and capability. As reported by TechCrunch, the executive was candid regarding their operational evaluations: "I think ChatGPT is better now. We publish a public benchmark as well, and consistently, Sol and Astra perform better on building vibe-coded applications from scratch. It was a decision based on both cost, partnership, and performance. And for all these reasons, we just felt like they were a better partner for now."
A formal partnership announced with OpenAI in August had already set the transition in motion. Because Replit executes code generation, terminal environments, and live hosting directly within the browser, the inference cost per generated token directly dictates unit economics. OpenAI's newer models delivered better reliability when assembling multi-file architectures while offering pricing structures that proved more sustainable for Replit's agentic loops than Anthropic's enterprise tiers.
Partner friction and shifting runtimes
The public announcement sparked immediate friction behind the scenes. Insiders familiar with the relationship between Anthropic and Replit pushed back against Masad's characterization, indicating that commercial agreements remain technically active. As late as September, technical write-ups and release notes for Replit Agent still highlighted Anthropic's Claude Fable 5 and 5.1 as active engines within production environments.
Nevertheless, Masad's remarks reflect an operational reality where traffic is systematically redirected toward OpenAI endpoints. The public acknowledgement signals that any legacy agreements with Anthropic are unlikely to see renewal once current contract terms conclude.
Toward the self-driving company
The model switch coincides with Replit's broader organizational redesign around agentic workflows. Outlining his vision during discussions covered by TechCrunch, Masad detailed the concept of a "self-driving company"—an enterprise where routine operational steps are managed by interconnected AI agents while engineers concentrate on taste, architectural balance, and judgment.
Within Replit itself, incoming customer support tickets are first processed by an autonomous agent that queries GitHub issues, examines source repositories, and inspects Linear boards before human staff step in. In this paradigm, raw model competence and cost efficiency take precedence over brand loyalty, turning model selection into a dynamic, performance-driven commodity.
Mocchi's take
The split between Replit and Anthropic offers clear validation that vendor lock-in around a single frontier model is unsustainable for modern software organizations. For engineering teams and businesses building agentic pipelines or bespoke applications, the lesson is clear: system architectures must remain strictly model-agnostic, routing individual prompts to whichever model delivers the best cost-to-performance ratio for that specific task. While rapid vibe coding drastically accelerates prototyping, human judgment regarding security boundaries, architectural resilience, and system integration remains strictly non-negotiable.