Replit Splits With Anthropic as CEO Masad Pitches the ‘Self-Driving Company’

The on-stage bombshell

Replit is no longer working with Anthropic. That was the headline disclosure from Amjad Masad, the coding platform's chief executive and co-founder, speaking on Thursday at TechCrunch Disrupt in San Francisco — and it instantly became one of the most talked-about moments of the conference.

Replit, the browser-based development environment that has become a favourite of startup founders and developers, announced a partnership with OpenAI back in August. But nothing in that announcement suggested its relationship with Anthropic was finished. Under the earlier arrangement, Replit powered its platform with Anthropic's Claude models — a tie-up that was credited with helping the company grow its revenue significantly.

Masad said the decision had been taken for "a number of different reasons, including cost". Asked to compare the two labs directly, he did not hedge: "I think ChatGPT is better now," he said, noting that Replit maintains an internal benchmark to assess model performance.

A split the paperwork may not recognise

There is a complication, however. People familiar with the Anthropic–Replit relationship have disputed Masad's version of events. One insider, speaking on condition of anonymity, insisted the partnership was still in place — flatly contradicting the chief executive's on-stage remarks.

The public record leans awkward, too. A Replit blog post published in late September cited Anthropic's Claude Fable 5 and Claude Fable 5.1 as models running in Replit Agent's production environment. Around the same time, one of Replit's own product employees was posting on X about Anthropic's flagship Opus 5.5 model, enthusiastically promoting its capabilities.

TechCrunch said it had contacted Replit for clarification and would update its reporting. Until that arrives, the industry is left with an odd spectacle: a chief executive announcing the end of a partnership that people close to it say never ended.

Cost, benchmarks and the economics of vibe coding

Whatever the contractual fine print turns out to be, Masad's stated reasoning is revealing. He framed the move around cost, the quality of the partnership and performance, measured against both an internal benchmark and a public one. OpenAI's Sol and Astra models, he claimed, consistently do better at building "vibe-coded" applications from scratch.

That benchmark claim matters for reasons well beyond bragging rights. Coding platforms consume model API calls at enormous scale, and inference costs feed directly into unit economics. For a platform serving everyone from weekend hobbyists to enterprise engineering teams, the choice of underlying model is not a marginal preference — it is a margin decision.

The performance argument also shows where the AI coding race is now being fought. "Vibe coding" — describing software in plain language and letting the model build it — has become the signature workflow of the category, and labs are competing specifically on who can generate working applications from a standing start. A public benchmark, whatever its methodology, is itself a competitive weapon.

Enterprise is becoming Replit's real business

Masad also used the Disrupt stage to sketch a striking change in Replit's revenue mix. Enterprise contracts made up barely 2 per cent of revenue last year; this year they account for roughly 15 per cent, and more than half of all revenue now comes from B2B customers. He declined to disclose annual recurring revenue.

The enterprise tilt helps explain the partnership calculus. Business buyers negotiate hard on price, demand reliability and scrutinise the models behind the tools they deploy. A partner offering stronger economics and better performance on the precise workload — generating apps from prompts — becomes the rational choice, whatever the history. It is the same enterprise-AI spending wave now dominating the markets, where investors are concentrating on the platforms actually capturing corporate budgets.

The "self-driving company"

The partnership drama shared the stage with Masad's bigger pitch: a vision he calls the "self-driving company". He does not mean robotaxis. He means firms in which AI agents absorb the grunt work while humans become more valuable for their judgement and taste.

"The reality is, all of our jobs — perhaps everyone in this audience, especially — changed over the past year," he told the Disrupt audience, adding that much of the world has yet to grasp the scale of the change heading its way.

Replit, he said, had spent the past year reorganising its own operations around agents after they became markedly more reliable earlier this year. Support tickets are now triaged by an agent first: it digs through GitHub issues, the codebase and Linear tickets, so that by the time a human sees the request, the relevant context is already assembled. When that person decides a patch is warranted, deploying a fix agent takes a single click.

The human, Masad insisted, is not being engineered out of the loop. The work agents do still passes through people: "agents did most of the work, but crucially, there's human judgment in the middle... You need human judgment; you need taste."

The result, he argued, is transformation rather than displacement — more time deciding what to build, more whiteboarding, more architecture, and work that is less boring and more creative. He closed by calling for Silicon Valley to offer a more optimistic public story about AI and work.

The whole industry is renegotiating

Replit's manoeuvring is part of a far larger pattern: 2026 is becoming the year the AI industry's partnerships got rewritten. In April, Microsoft and OpenAI overhauled their landmark alliance. Microsoft kept its primary cloud partner status and rights to OpenAI's models and products through 2032, but the licence turned non-exclusive, OpenAI won freedom to serve customers from any cloud, and OpenAI's revenue share to Microsoft was capped at roughly 20 per cent through 2030, with Microsoft dropping its own revenue share in return.

Anthropic's filings, meanwhile, revealed the staggering scale — and rigidity — of its infrastructure commitments: roughly $518 billion of planned buildout spending, including a $31.4 billion commitment to Microsoft running from November 2026 to May 2033 that cannot be cancelled except for material breach, alongside major arrangements with Google, Amazon, xAI and AMD. The company warned that the tech giants backing it act simultaneously as investors, customers, cloud providers, distributors and competitors, with incentives that "may not be fully aligned" with its interests.

Add OpenAI's reported $50 billion investment from Amazon, the $500 billion Stargate infrastructure programme and a string of xAI compute deals, and the pattern is unmistakable: every lab is shopping for leverage, and every platform — Replit included — is re-rating its suppliers. Partnerships that looked permanent two years ago now look provisional.

What it means for the AI coding race

For developers, the immediate question is practical: which models actually sit behind Replit's prompts, and how good are they at generating working software? Masad's benchmark claims are self-interested, but they point to where competition is heading — public, workload-specific shootouts over who builds the best app from a blank page.

For the labs, Replit is both a prestige logo and a distribution channel. Losing it — if the split is as final as Masad suggests — would sting Anthropic precisely as Claude's developer mindshare faces its toughest test. For OpenAI, adding Replit to its roster consolidates a platform strategy built on ubiquity, from consumer chatbots to enterprise coding tools.

The deeper lesson is structural. In 2024, cosy lab–platform pairings were the norm; in 2026, platforms are behaving like procurement departments — benchmarking, haggling over cost and switching suppliers. Expect more chief executives to take to stages with news of who they are no longer working with.

FAQs

Has the Replit–Anthropic split actually been confirmed?

Not unambiguously. Masad announced it on stage at TechCrunch Disrupt, but an insider familiar with the relationship told TechCrunch the partnership was still in place, and Replit's own recent output — a late-September blog post citing Claude Fable 5 and 5.1 in production, plus an employee promoting Opus 5.5 — suggests Claude models were in use weeks ago. TechCrunch has asked Replit to clarify; until it does, treat the split as announced but not fully documented.

Why would Replit choose OpenAI over Anthropic?

Masad cited three factors: cost, the quality of the partnership and performance. Replit runs internal and public benchmarks, and he claimed OpenAI's Sol and Astra models do better at building vibe-coded apps from scratch. For a platform consuming API calls at massive scale, even small differences in price or performance translate directly into margins — which is why model choice has become a procurement decision, not a loyalty one.

What is the ‘self-driving company’?

Masad's term for an organisation in which AI agents handle routine execution while humans supply judgement and taste. At Replit, agents now triage support tickets by searching GitHub issues, code and Linear boards before a human sees them, and a single click can deploy a fix agent once a person approves a patch. The company reorganised around this pattern over the past year, after agents became markedly more reliable.

Is this part of a wider industry trend?

Yes. In April 2026, Microsoft and OpenAI rewrote their alliance to be non-exclusive and multicloud, with capped revenue sharing and IP rights through 2032. Anthropic's filings disclosed roughly $518 billion in infrastructure commitments, much of it non-cancellable, while warning that its Big Tech backers are simultaneously investors, suppliers, distributors and competitors. Labs and platforms are renegotiating constantly as costs, capabilities and leverage shift.

What should Replit users watch for?

Which models actually serve their prompts in the coming weeks, and whether Replit issues a formal clarification about Anthropic. Longer term, the company's enterprise push — from 2 per cent to 15 per cent of revenue in a year, with B2B now over half the total — suggests more business-focused features, pricing and controls are on the way.