Replit CEO Says It Has Dropped Anthropic as OpenAI Partnership Deepens

Replit is no longer working with Anthropic, CEO and co-founder Amjad Masad said Thursday at TechCrunch Disrupt, signaling a notable shift in the AI coding platform’s model strategy as it leans more heavily on OpenAI. The change matters because Replit sits at the center of the fast-growing “vibe coding” market, where model quality, price, and reliability can directly shape developer adoption and revenue.

Masad said the decision came down to a mix of factors, including cost, product performance, and the evolving partnership landscape. His comments came months after Replit publicly announced a collaboration with OpenAI, and they arrive as the company’s business mix changes rapidly, with enterprise and B2B accounts now representing a much larger share of sales than they did a year ago.

But the announcement also raised questions. People familiar with Replit’s Anthropic relationship pushed back on the idea that the partnership was fully over, and a recent Replit blog post still referenced Anthropic models in production. Replit did not immediately clarify the situation, leaving room for uncertainty about whether the company has ended a direct commercial relationship, reduced its use of Claude, or simply shifted which models power the platform most prominently.

What Masad said onstage at Disrupt

At TechCrunch Disrupt, Masad framed the move as a practical business decision rather than a dramatic break. He said Replit evaluates model providers through internal and public benchmarks and concluded that OpenAI’s latest systems were a stronger fit for the products Replit is building.

According to Masad, the company looks at both general model performance and how well the systems support app creation from a blank page. He said Replit’s own benchmark suggests its “Sol” and “Astra” systems perform better for building applications from scratch in the company’s specific environment.

“It was a decision based on both cost, partnership, and performance,” Masad said, explaining why Replit favored OpenAI.

His remarks echoed a broader pattern in the AI industry, where application companies often swap model providers as pricing, feature quality, latency, and commercial terms evolve. In practice, many AI products do not rely on a single foundation model forever, especially when their user experience depends on fast iteration and margins that can withstand heavy inference costs.

How did Replit’s model strategy change?

Replit’s AI stack appears to have shifted from a stronger reliance on Anthropic’s Claude models toward a closer relationship with OpenAI. That does not necessarily mean every Anthropic system has been removed from the product, but it does suggest the center of gravity is moving.

The company announced a partnership with OpenAI in August, a move that signaled a deeper alignment with the ChatGPT maker. Until Masad’s Disrupt appearance, there had been no public indication that Anthropic was being displaced, which made the announcement notable for observers tracking the competitive dynamics among model providers.

Replit’s core offering is a browser-based development environment that allows users to create and deploy applications entirely online. Because the product is designed to help people code quickly, often with natural-language prompts, model selection is not just a technical detail; it is part of the user experience and business model.

Why model choice matters for coding platforms

For an AI coding company, the underlying model can determine how well the product handles planning, code generation, debugging, and iterative editing. A stronger model can reduce user frustration and boost completion rates, while a cheaper model can make the service economically viable at scale.

That trade-off is especially important in “vibe coding,” a term used for AI-assisted app creation that emphasizes rapid prototyping and conversational coding workflows. In this market, users often expect near-instant responses and code that works on the first or second try. If the experience slips, they may switch tools quickly.

Masad’s comments suggest Replit is optimizing for a blend of quality and economics rather than loyalty to one provider. That stance reflects a maturing AI application market in which customers care less about which lab powers a product and more about whether the product works reliably, securely, and at a reasonable cost.

Did Replit really end its Anthropic relationship?

That is the key question, and the answer is not fully clear from the public record. While Masad said Replit is “no longer working with Anthropic,” a person familiar with the partnership told TechCrunch the relationship was still in place, creating an apparent contradiction between the executive’s remarks and outside knowledge of the deal.

There is also documentary evidence suggesting Anthropic models were still in use recently. A Replit blog post published in September cited Claude Fable 5 and Claude Fable 5.1 as models running in Replit Agent’s production environment, which indicates Anthropic technology may still have been part of the platform after the OpenAI partnership was announced.

The most likely explanation is that the company is in the middle of a transition, rather than a clean, immediate break. In the AI industry, partner changes can be messy: one model family may remain active for specific workflows even as a newer provider takes over the headline role.

TechCrunch said it had reached out to Replit for clarification and would update its reporting if the company responded. Until then, the safest interpretation is that Replit’s model lineup is evolving, but the exact boundaries of the Anthropic relationship remain unresolved.

Replit’s revenue mix is changing fast

Masad also used his Disrupt appearance to offer a glimpse into Replit’s commercial growth. He said enterprise and business customers now account for a much larger share of the company’s revenue than they did last year, underscoring Replit’s shift from a developer-friendly startup tool toward a broader software platform with corporate traction.

He stopped short of revealing annual recurring revenue, but said enterprise made up only a small slice of revenue last year and now contributes meaningfully more. More broadly, he said more than half of Replit’s revenue comes from B2B customers.

That is a significant signal for investors and competitors alike. AI coding platforms often start with individual users and small teams, but the path to durable growth typically runs through organizations willing to pay for security, governance, collaboration, and support.

How Replit is thinking about AI safety

Masad also used the stage to make a broader argument about the conditions AI agents need in order to work safely in real-world settings. His framing went beyond model quality and into infrastructure, security, and controls.

He compared agents to human workers, saying they need an environment that lets them function effectively. In his view, that environment includes orchestration, recovery systems, sandboxing, networking, identity management, and authorization. Without those layers, he suggested, companies should not expect reliable or secure deployment.

Masad argued that AI systems need a proper “habitat” with safety, security, and operational guardrails before they can be trusted in production.

This is an increasingly common perspective among AI platform leaders. As coding agents become more capable, the challenge shifts from whether they can generate code to whether they can operate safely inside live software systems without breaking permissions, leaking data, or making costly mistakes.

Why safety is becoming a product feature

For Replit, safety is not just a compliance issue. It is part of the value proposition. Customers want AI tools that can help them build software faster without introducing hidden risk, especially if those tools are allowed to access codebases, cloud resources, and deployment pipelines.

That means companies such as Replit must balance speed with controls. Too many guardrails can make an AI assistant feel clunky. Too few can make it dangerous. The companies that solve that trade-off may have the best shot at serving enterprise customers at scale.

What the shift says about the AI model wars

Replit’s apparent move away from Anthropic, and toward OpenAI, illustrates how fluid the AI supplier landscape has become. Startups that build on foundation models are increasingly acting like sophisticated procurement departments, comparing pricing, throughput, feature releases, and contractual terms across vendors.

That dynamic creates constant pressure on model labs. A strong technical lead can be temporary if another provider offers better economics or a more favorable product bundle. For model companies, the challenge is no longer just building the best model; it is keeping customers locked in as the market matures.

Replit is especially worth watching because it sits at the intersection of consumer-like developer adoption and enterprise software sales. A single product decision can influence not only user behavior but also how much revenue the company can extract from different segments.

For Anthropic, any loss of visible platform partners would be notable, particularly in the coding category, where Claude has been widely respected for software-related tasks. For OpenAI, winning more of Replit’s workload would further strengthen its position in the developer tooling ecosystem.

Timeline: how the story developed

The chronology helps explain why Masad’s comments caught attention.

What happens next?

The immediate next step is clarification. Replit has not publicly laid out whether the Anthropic relationship is fully terminated, partially reduced, or simply no longer central to its platform strategy. A company response could clear up whether Masad was describing a product shift, a business decision, or a formal contract change.

Until then, the safest conclusion is that Replit is reweighting its dependence on model providers as it grows. The company appears to be betting that a combination of OpenAI’s performance, commercial terms, and strategic alignment better supports its next phase of expansion.

That shift is more than a vendor switch. It is a sign of where the AI application market is heading: toward flexible, benchmark-driven partnerships where model choice is constantly reassessed against cost, customer demand, and the realities of shipping software at scale.

For founders, developers, and enterprise buyers, the message is clear. In AI infrastructure, no partnership is permanent, and the best model today may not be the one powering the product tomorrow.

Key facts at a glance

• Replit CEO Amjad Masad said the company is no longer working with Anthropic.
• He said the decision was driven by cost, performance, and partnership factors.
• Replit announced an OpenAI partnership in August.
• Masad said enterprise revenue has grown from about 2% to around 15% year over year.
• More than half of Replit’s revenue now comes from B2B customers.
• Questions remain because outside sources say the Anthropic relationship may still exist.

If you want, I can also provide a shorter version of this story, a headline set, or a matching social media summary package.