Build · founder · 5 min read

Replit Says It Dropped Anthropic. Here's What Changes for Your Builds

Replit's CEO says the company no longer works with Anthropic and prefers OpenAI models. What is confirmed, what isn't, and what to check in your account.

Published October 17, 2026 ·

replit anthropic openai models vendor risk

On October 16, Replit CEO Amjad Masad told a TechCrunch Disrupt audience that his company is no longer working with Anthropic. If you build on Replit, the headline sounds alarming. The reality so far is murkier, and for most people the right response is a short checklist, not a migration.

What was actually said

According to TechCrunch and Unite.ai, Masad answered a question about the OpenAI partnership by confirming the Anthropic relationship is over. His stated reasons were cost, partnership terms and performance. He said OpenAI’s current models do better at building vibe-coded apps from scratch, citing an internal benchmark and a public one.

He also said Replit was one of Anthropic’s top five customers, and that enterprise has grown from about 2% of revenue last year to about 15% this year, with more than half of revenue now coming from businesses.

What is not settled

Three things don’t line up, and you should hold the headline loosely.

• The product still lists Claude. Replit’s model selector documentation still shows several Claude models as options in Max mode, per Unite.ai’s reading.
• A source disagrees. TechCrunch reports an unnamed person familiar with the arrangement who says the Anthropic partnership is still in place.
• Late September looked different. A Replit blog post then described Claude models running in Agent’s production environment.

Neither Replit nor Anthropic has given a clarifying statement in the coverage we found. “No longer working with” could mean an ended commercial deal, a shift in default models, or loose phrasing on stage. We don’t know yet, and we’d rather say so than guess.

Why Replit would do this

The direction isn’t a surprise. Replit launched Free Mode on OpenAI’s GPT-5.6 Luna in August, which made everyday Agent chat stop consuming credits. Its September engineering post reported the Agent running on GPT-6 Astra as the core loop in Max mode, at roughly $2.11 per task on one coding benchmark. Those are Replit’s own numbers, so treat them as a vendor claim, not an independent test.

The business logic is simple. A builder that resells model access lives on the gap between what the model costs and what you pay. Whoever offers the better price per successful task wins that margin, and Replit has clearly picked its preferred supplier.

What this means for you

If you’re a non-technical founder on the $20 Core plan, most of your day-to-day work already runs through Free Mode and automatic routing. You likely won’t notice anything.

The people who may notice are those who deliberately picked a Claude model for complex builds. Output style, how the agent handles ambiguity, and how many credits a task burns can all shift when the default model changes. That is the real risk: not that Replit breaks, but that your project quietly behaves differently.

This is the same lesson as in our guide on which model powers your vibe coding tool. The tool is a wrapper. The model underneath can change without you signing anything.

A short checklist

1. Open the model selector. Note which model each mode (Free, Power, Max) uses on your account today. Screenshot it.
2. Rerun one familiar task. Pick a feature you’ve built before and compare quality and credits spent against your past results.
3. Check your credit usage page for a week. If Max mode costs more or less than before, adjust which mode you default to.
4. Keep your code portable. Make sure the project is synced to your own GitHub repository, so a pricing or model change never traps you.
5. Watch for an official statement. Both companies will probably clarify, and the answer could change the advice here.

The bigger picture

The AI builders are becoming model-agnostic storefronts. Cursor, Replit, Lovable and others route requests to whichever vendor is cheapest or best this quarter, and vendors now court them with exclusive deals. For you, that means more price competition, and also less stability. Your favorite tool’s personality is partly borrowed from a supplier that can be swapped.

Our advice stays consistent: choose tools for the outcome they deliver on your project, keep your code in a repository you control, and re-test when the vendor changes anything major. For a wider view of how this fits with pricing shifts, see the model price war.