Nobody got an email.
That is the part I keep coming back to. Replit, the AI coding platform that turned “describe it and we’ll build it” into something millions of people actually use, swapped the default model behind its coding agent from Anthropic’s Claude to OpenAI’s. The announcement came from its CEO on a conference stage, not from a changelog entry or a customer notice. If you build on Replit and you leave your settings alone, the builder behind your app got a new brain this month. Nobody asked you first.
That sounds like a product story. It is actually an economics story, and it has a lesson in it for anyone who ships software on top of somebody else’s platform.
What actually happened
On stage at TechCrunch Disrupt, Replit CEO and co-founder Amjad Masad was asked whether the OpenAI partnership the company announced in August meant it had stopped working with Anthropic. He did not dance around the question. “Yeah. I mean, yeah,” he answered, according to unite.ai’s account of the session, and told TechCrunch: “I think ChatGPT is better now.”
The stated reasons were cost, partnership, and performance. Masad said OpenAI’s GPT-6 Sol and GPT-6 Astra “consistently” perform better on building applications from scratch. He also disclosed that Replit had recently been a top-five Anthropic customer, which is what makes this more than a routine vendor shuffle: one of Anthropic’s largest pipeline customers just redirected its default traffic somewhere else.
Then came the walk-back. Two days later Masad posted a clarification that softened the framing without reversing the decision: “Replit remains model agnostic. Our users can choose Anthropic models as well as many other supported models. What I meant at TC Disrupt is that OpenAI is now our preferred partner powering Agent based on performance, cost, and strength of our partnership.”
Notice what changed and what did not. The default changed. The catalog did not. Replit’s own documentation still lists seven Claude models, from Sonnet 5 and Sonnet 4.6 in Power Mode to Fable 5.1, Opus 5.5, Opus 5, Opus 5 Fast and Opus 4.8 in Max Mode. Claude is still there. You just have to go looking for it, and one tier of users cannot look at all: Free Mode builders, Replit’s docs say, “cannot use the Model selector or choose a model manually.” Choice, in other words, is a paid feature.
TechCrunch also reported that a person familiar with the partnership said it was still in place, and that a late-September Replit blog post still named Claude Fable 5 and Fable 5.1 as running in Replit Agent’s production environment. So the business question, whether the partnership is truly over, stays open. The product question does not: the default moved, and the default is where the volume goes.
Why the default model is worth more than the best one
Here is the economics most of the coverage skipped past. Masad said enterprise contracts were “like 2% of revenue last year” and roughly 15% this year, and that “more than 50% of our revenue is B2B.” He expects enterprise and government business to pass half of revenue in 2027.
Enterprise buyers are the ones who read documentation. Everyone else takes the default. And TechCrunch’s reading of the switch is blunt: most customers stick with the default rather than selecting a model themselves, so moving it could significantly reduce Replit’s spend with Anthropic on its own. Flip the default and you do not have to fire anyone or renegotiate anything. You just stop asking.
Replit’s own roadmap makes the point better than any analyst could. In late August the company shipped “intelligent model routing,” and made Auto mode the default Agent mode for Enterprise Workspaces. Admins can restrict which models a workspace can reach, and can switch Auto mode off, but only by picking a specific model from the list the platform allows. The platform is telling you, in product language: we will pick the model for you now, and you may choose from the menu we hand you.
This is the same pressure Anthropic has been navigating from the other side. The company has spent the year positioning itself as the transparent alternative, opening its models up to outside audits whose results it cannot veto, and this month committing 5% of its shares to a public benefit fund. Positioning matters for the narrative. It does not route traffic. Defaults do.
The contract gap nobody reads
If you run a team that builds on Replit, the uncomfortable question is whether anyone told you. The answer, according to enterprise vendor-risk analyst Rajesh Beri, is probably no.
OpenAI was already on Replit’s subprocessor list, last updated in August 2026, sitting right next to Anthropic and Google. Replit’s data processing agreement, dated April 1, 2026, gives the company “a general authorization” to engage subprocessors and sets no advance-notice period. Your contractual right to object attaches to “a change of a Subprocessor.” Shifting the default between two vendors who were both already on the list is, on the paper you signed, not a change of subprocessor. It is a change of routing.
Which means the privacy assessment your team signed off on, the one that said prompts and code mostly go to Anthropic, may no longer describe where your data actually flows. That is the gap, and it belongs on the checklist next to the questions you already ask about uptime and pricing. It is also not unique to Replit. Platforms quietly rewriting the terms that govern your data has become a genre of its own, right down to who gets to train on your website.
This is not a one-off
Look at the same stretch of calendar. On October 22, Uber announced it was rolling ChatGPT Work out to more than 30,000 employees worldwide, its first company-wide deployment of an agentic platform, with GPT-6.1 Sol deployed as the default capability inside it. Atlassian’s OpenAI deal moved Rovo routing without giving administrators a provider choice. IBM’s self-hosted Bob swapped out Claude, Mistral and Granite in favor of its own Nemotron and Laguna models.
Different companies, identical shape. The model behind the product stops being a decision you make and starts being a decision that was made for you. Sometimes it arrives as a default. Sometimes as a routing layer you cannot switch off. Sometimes as the only option in the tier you can afford.
What I would tell a developer building on someone else’s platform
I run AI agents every working day, and I have written about the skills that actually matter when you use AI tools: prompt discipline, verification, and knowing when to reach for a different model. This story is the argument for all three.
Pin your models. Not prefer, pin. If your build depends on a specific model behaving a specific way, write that into configuration rather than leaving it in a settings menu that a platform can change on a Tuesday. Keep a private evaluation set, a handful of tasks your system has to get right, and re-run it every time the platform announces anything at all. The cost of that habit is trivial. The cost of finding out from a user that your app broke is not.
Own your distribution. Replit is a genuinely good place to prototype. It is a risky place to keep the only copy of your product. The lesson is not “do not use platforms.” It is “do not let a platform be the only door to your customer.”
And if concentration worries you, a handful of providers routing an enormous share of the world’s software-building traffic, the honest answer is that running local AI for daily tasks is no longer a hobbyist’s project. Running a serious model on your own hardware is a real option for the parts of your stack you cannot afford to have rerouted underneath you.
The bottom line
Nobody got an email, and that is the tell. The model behind your tool changed because a platform decided it should, not because you did. Anthropic and OpenAI are both excellent, and the benchmarks will keep bouncing between them for years. The question was never which one is better. It is who gets to decide, and whether you would even notice if the answer changed.
I manage systems where a silent change in the plumbing is exactly the kind of thing that shows up as an incident three months later. So I would notice. That is why I pin.