What happened
Replit sells a browser-based platform where people build and deploy apps, including with an AI agent. According to TechCrunch, Masad said onstage at Disrupt on Thursday that Replit is no longer working with Anthropic. He said the decision was based on "cost, partnership, and performance," and that OpenAI's models "perform better on building vibe-coded applications from scratch" on Replit's internal benchmark and a public one. TechCrunch published that on October 16, then updated it on October 17 with more context. After the talk, Masad said customers can still use Anthropic's models on Replit, and that OpenAI is now the default. He also said Replit "remains model agnostic" and that OpenAI is its "preferred partner for powering Replit Agent." The account is not clean. TechCrunch reports that a person familiar with the relationship rejected the on-stage framing and said the partnership is still in place. It also found a late September Replit blog post naming Claude models as running in the Agent's production environment. Replit announced a partnership with OpenAI in August. TechCrunch's piece includes no comment from Anthropic or OpenAI. No benchmark scores, prices, or savings were published. So the confirmed change is the default and the preferred partner. Whether the Anthropic relationship has ended is disputed.
Why it matters for business owners
You may never use Replit. The mechanism applies to a lot of software you do pay for. Many AI products are a layer on top of someone else's model. The product has its own name, price, and interface, and a model from OpenAI, Anthropic, Google, or another lab doing the work underneath. The company selling the product chooses that model, and it can change the choice. TechCrunch makes the key point: most customers stay on the default rather than picking a model themselves. So a default change moves most of the product's usage to a different model, even though the menu still lists the old one. That can change the quality of the output, the style of the work, the speed, and the cost per task. It can also change where your data goes, because a different lab may sit behind the tool and the vendor's terms with that lab apply. Your contract is with the tool vendor, but your work is now handled by a company you may never have looked at.
What owners should not misunderstand
Do not read this as proof that one lab beat another. The performance claim is Replit's own, from its own benchmarks, and no scores were published. It is also about one job: building apps from scratch. It says little about drafting contracts, reading invoices, or answering customers. Do not treat a CEO's remark onstage as a settled fact. The first headline said Replit was no longer working with Anthropic. After clarification the story became a changed default. A source says the partnership continues. Wait for the company's own written statement before you change a plan. Do not assume a cheaper default means a cheaper bill for you. Masad cited cost, but no figures were given, and nothing says savings reach customers. A different model can also need more or fewer attempts per task, which changes your cost per finished result. Do not assume "model agnostic" means nothing changes for you. It means you can choose. It does not mean you will choose, or that the default behaves the same as before. And do not conclude that you should switch tools. A supplier change is a reason to check, not a reason to panic.
The operational lesson
A model is a dependency, and you can only manage dependencies you can see. Most owners can name the tools they pay for. Fewer can say which model sits behind each one, who picks it, and how they would find out if it changed. The Replit story is a public example of something that normally happens quietly. Two things protect you. The first is a written record of which tools touch which work, and what that work needs to look like when it is right. The second is a small set of real tasks you can rerun whenever a vendor changes something. Without them, you notice a change when output gets worse, or the bill moves, or a customer complains. The same logic applies to vendor claims. Benchmarks measure what the vendor chose to measure. Your own examples measure your own work.
What a serious business should do next
List the AI tools your team uses for real work. For each one, find out which model provider runs it, whether you can choose, and what the default is. If the vendor does not say, ask. Read how the vendor handles model changes. Does it announce them? Does it let you pin a model, or does it route automatically? Some products pick a model per task for you, which is convenient and also means the answer to "which model did this" can differ from one day to the next. Build a small test set. Pick three to five real tasks, with non-sensitive material, that represent what the tool does for you. Save what a good result looks like. When a vendor changes its model or default, rerun them. Check the data terms of the new supplier. A different lab can mean different retention and training terms. Your agreement with the tool vendor should say where your data goes. Keep your work portable. Keep prompts, templates, and examples in your own files, not only inside one product, so a switch costs days and not months. If your team builds apps with a tool like this, remember that code written under one default may need review when the default changes. Review rules should not depend on which model wrote the code. Check current plan and model pages before deciding. Defaults and pricing in this space change often.
The Atlacis view
Atlacis helps owners slow down before an AI decision, map the workflow, and match the tool to the job. A tool is only as stable as the supplier behind it, and the supplier can change without you being asked. If you are unsure which of your tools depend on which models, or what you would test if one of them changed, that is a useful thing to map before it becomes urgent.
The short version
- Replit's CEO said at TechCrunch Disrupt on October 15, 2026 that Replit is no longer working with Anthropic. TechCrunch later reported OpenAI is the default for Replit Agent and customers can still choose Claude.
- A source familiar with the relationship told TechCrunch the Anthropic partnership is still in place. The ending of the relationship is disputed.
- No benchmark scores, prices, or savings were published. The performance and cost claims are Replit's own.
- Most users stay on the default model, so a default change moves most usage even when the old model is still listed.
- Know which model sits behind each AI tool, keep a small test set of your own tasks, and rerun it when a vendor changes something.