Michael H

Posted on Oct 20 Originally published at agent5.news

Replit and Anthropic: What Really Happened and Why It Matters

#ai #technology #aicodingtools #modelpartnerships

Originally published on agent5.news.

When Replit CEO Amjad Masad took the stage at TechCrunch Disrupt on October 16, 2026, he said something that caught the AI industry off guard: his company was no longer working with Anthropic. Within hours, insiders disputed the claim, and Masad himself softened his words. What looked like a clean breakup turned into something far more revealing about the current state of the AI coding market.

What Masad Actually Said Onstage

The exchange started with a direct question: did the OpenAI partnership Replit announced in August 2026 (the exact date of which has not been independently confirmed) mean the company was no longer with Anthropic? Masad's answer was blunt. He confirmed the split and attributed the decision to three factors: cost, performance, and the strength of the partnership with OpenAI. "I think ChatGPT is better now," he said, citing Replit's internal benchmark and a public one the company publishes, adding that OpenAI's GPT-6 Sol and GPT-6 Astra models consistently performed better at building vibe-coded applications from scratch.

The statement was notable because it was the first time Replit had publicly acknowledged any shift away from Anthropic. When Replit had announced its OpenAI partnership in August 2026, the company, according to TechCrunch, framed it as an expansion and gave no indication that Anthropic was being sidelined.

The Immediate Pushback

The backlash came fast. A source familiar with the Replit-Anthropic relationship told TechCrunch that the partnership was still in place. Separately, Replit's own documentation contradicted the clean-break narrative: a post on Replit's engineering blog, reported by secondary sources as dated September 29, 2026, documented Claude Fable 5 running in Replit Agent production in August 2026, and Claude Fable 5.1 alongside OpenAI's GPT-6 Astra in September 2026, just weeks before Masad's TechCrunch appearance.

By the weekend after TechCrunch Disrupt, Masad had walked back his sharpest comments. According to TechCrunch, he clarified that Replit remains model agnostic, that users can choose Anthropic's Claude or other supported models, and that OpenAI is now the preferred partner for powering Replit Agent based on performance, economics, and the strength of the partnership.

In other words, the reality was more nuanced than either a breakup or a status quo. Anthropic had not been cut out of the platform entirely, but it had been demoted from default partner to one option among many.

Why the Default Model Matters So Much

The distinction between "default" and "available" is not a minor technical footnote. Most users of any software product never change the default settings. For Replit, which operates a cloud-based development platform where users build and deploy apps directly inside a browser, the default model is the model that most users will end up using for every project, every prompt, and every API call. Shifting the default away from Claude and toward OpenAI's models is, in practical terms, a major reduction in Replit's spend with Anthropic, even if Anthropic's models technically remain accessible.

That makes the default model decision one of the most commercially important levers in a platform business. For Anthropic, losing the default slot at a top-five customer is a meaningful revenue event, regardless of what the formal contract says.

The History Behind the Split

The Replit-Anthropic relationship was not a minor vendor arrangement. Claude models had powered Replit's platform for years, and the collaboration was publicly celebrated by both companies. Anthropic's customer case study page for Replit, which has been cited in published sources, described how Replit used Claude on Google Cloud's Vertex AI to help anyone build and deploy software applications without coding experience. Google Cloud and Anthropic case study sources referenced Replit using Claude models on Vertex AI to power Replit Agent, and through 2025 Replit publicly adopted newer Claude-family models as they became available.

Masad acknowledged the scale of that relationship onstage, saying, according to multiple secondary sources, that Replit had recently been one of Anthropic's top five customers. That scale also illustrates just how quickly the model landscape can shift when a platform the size of Replit decides to prioritize a different provider.

What Replit's Partnership Web Actually Looks Like

One of the underreported dimensions of this story is how many partners Replit was already managing simultaneously. In December 2025, Google Cloud announced a multi-year partnership with Replit, under which Google would remain Replit's primary cloud provider, Replit would expand its use of Google Cloud services, add more of Google's models onto its platform, and support AI coding use cases for enterprise customers. In July 2025, Replit also partnered with Microsoft to make its platform available through the Azure Marketplace.

So at the moment Masad made his TechCrunch comments, Replit had active relationships with Google Cloud as its primary infrastructure provider, Microsoft for enterprise distribution, and now OpenAI as its preferred model partner. Anthropic was the odd one out, demoted from a central role to an optional one. Replit's own model selector, as of mid-October 2026, still listed multiple Claude variants alongside GPT and Gemini models, confirming the platform is multi-model rather than monoculture.

The Three Reasons Masad Gave (and What They Mean)

Masad summarized the move with three words: cost, partnership, and performance. Each one is worth unpacking.

On cost: for a developer platform, the price paid per API token is directly tied to unit economics. Replit generates revenue by charging users for compute cycles and credits, and a meaningful slice of that cost flows to whatever model provider is running in the background. If OpenAI was offering more favorable pricing, switching the default would improve margins on every single project run on the platform. OpenAI has been pricing its models competitively to win platform partnerships, and that strategy appears to be working.

On performance: Masad said Replit uses both an internal benchmark and a public one, and that on both, OpenAI's GPT-6 Sol and GPT-6 Astra models outperformed Claude for the specific task of building vibe-coded applications from scratch. It is worth noting that this benchmark covers one particular workload. Building a new app from a natural-language description is different from editing a large existing codebase, debugging, or writing security-sensitive code. Different models have different strengths, and Replit's benchmark reflects Replit's primary use case.

On partnership: this is the most opaque of the three reasons, but in practical terms it likely refers to the level of commercial support, co-marketing, and integration assistance that OpenAI was prepared to offer. Model providers compete not just on technical performance but on the support infrastructure they build around their biggest platform customers.

Replit's Business Context: A Company at Speed

The timing of this shift is inseparable from Replit's extraordinary growth trajectory. In March 2026, Replit raised a $400 million Series D led by Georgian at a $9 billion valuation, three times its valuation from just six months prior. Sacra estimated Replit reached roughly $525 million in annualized revenue by April 2026, up from an estimated $300 million at the end of 2025. The company itself has said it is targeting $1 billion in annual recurring revenue by end of 2026.

Enterprise has become a growing driver: Masad said onstage that enterprise contracts made up about 15% of revenue, up from 2% the year before, and that he expects enterprise and government customers to account for more than half of Replit's revenue in 2027. A company growing that fast has acute sensitivity to model costs and vendor terms. At scale, even a modest per-token cost advantage compounds into significant savings, and the calculus around which model to default to becomes a financial decision as much as a technical one.

Replit also has more than 50 million users, and the company has said users from 85% of Fortune 500 companies are on the platform. At that scale, the model default setting is not a preference, it is a policy with material consequences.

What This Signals for AI Coding Tools Broadly

The Replit situation is a window into a dynamic that is playing out across the AI coding landscape: model providers are not just competing for developer attention, they are competing for the default slot on high-volume platforms. Cursor, Lovable, and other tools in the vibe-coding space each have their own model partner strategies, and those strategies are increasingly subject to renegotiation as costs shift, performance rankings change, and commercial relationships evolve.

For users, the practical implication is that the underlying model powering a favorite tool can change without obvious announcement. The interface looks the same, the product name stays the same, but the model generating the code underneath may be entirely different from one month to the next. This is not necessarily a problem, but it does mean that users who care about which model is running their code need to check the settings actively, not just trust the default.

For model providers like Anthropic, the episode underscores the risk of platform concentration. Being a top-five customer's primary model provider is a significant revenue position, but it is also a fragile one if the platform can re-route traffic to a competitor by changing a single default setting.

What to Watch Next

Several concrete developments are worth tracking in the weeks and months ahead. Replit has not issued a formal joint statement with either Anthropic or OpenAI clarifying the exact status of its agreements. A formal clarification from any of the three parties would resolve the dispute about whether the Anthropic contract is still active.

Replit's model selector documentation is also worth monitoring. As of mid-October 2026, Claude models were still listed as selectable options in both Power and Max Modes. Any changes to that list would signal whether the relationship is truly winding down or stabilizing at the non-default level.

The company's revenue trajectory toward its stated $1 billion annual target by end of 2026 will also matter: hitting or missing that goal will influence which commercial deals Replit can negotiate from a position of strength. And the broader question of how OpenAI, Google, and Anthropic each continue to price and support their platform partnerships will shape which defaults developers encounter every time they open an AI coding tool.

DEV Community