Replit drops Anthropic as default, switches to OpenAI

Full analysis

Replit, once a top-five Anthropic customer, just made OpenAI's ChatGPT the default model behind Replit Agent, the tool that builds whole apps from a prompt. CEO Amjad Masad said the call came down to "cost, partnership, and performance." Anthropic's models stay available if you pick them by hand, but almost nobody does, so Anthropic's money from Replit is about to fall off a cliff.

This is easy to undo on paper and hard to undo in practice. Flipping a default is one config change. But Replit will now rebuild its prompts, guardrails, and error handling around OpenAI's behavior, and that work calcifies fast. What's actually being decided here is not "which model is smarter." It's what the going rate is for a big, visible, token-hungry coding customer, and who's willing to eat margin to get the logo. No deadline forced this. Masad chose the moment, on stage at Disrupt, which tells you something.

The Skeptic. Read the quote again. "Cost, partnership, and performance." Performance is third on the list, and "partnership" is a word that means whatever the person saying it wants it to mean. Everyone is turning this into "OpenAI wins the coding war." The cleaner explanation is that OpenAI wrote a better check for a trophy customer. Replit is the worst kind of account to serve: enormous token volume, price-sensitive, agentic loops that hammer rate limits, thin margin. Losing it might help Anthropic's unit economics, not hurt them. A design-partner logo is worth overpaying for, and OpenAI overpaid. That's business development, dressed in benchmark language.

The Compute Pragmatist. Replit's loop is multi-turn: the model writes code, runs it, reads the error, tries again. That burns output tokens like nothing else, and output tokens are where premium pricing bites hardest. If OpenAI offered a lower effective rate or a volume commitment Anthropic wouldn't match, the performance numbers only need to be a tie for the switch to make sense. That's the signal worth carrying out of this story. Big agentic buyers now have real leverage to squeeze the labs, because every lab wants reference deployments and the inference price floor keeps dropping. Training still costs a fortune. Serving a whale at a discount to win the logo compresses margin on both ends.

The Researcher. The benchmark point is narrow but real. "Vibe-coded" app generation, the quick-scaffold, build-from-scratch loop, is its own task. Rankings on it don't have to match HumanEval or SWE-bench, the standard coding tests the labs quote. Masad is reporting a live A/B test at a scale most of us can't run, so it's a credible data point. The awkward read for Anthropic: Claude's strengths, long-context reasoning and careful instruction-following, may not be what this loop rewards. Fast, iterative scaffolding is a different sport than holding a huge codebase in your head. One production test isn't a verdict. But it's not nothing.

The Safety Lens. Replit Agent builds full apps from a prompt with almost no human review, then users ship them. Swap the default model and you change the behavior for millions of auto-generated codebases at once: how credentials get handled, which dependencies get pulled, whether the database query is wide open to injection. Neither lab has published safety evals specifically for agentic code generation at scale. Nobody knows which default produces fewer insecure apps in the wild, because nobody's measuring it. The model swap isn't the risk. The risk is that the platform optimizes for build speed, and insecure code compounds quietly across every project built on whatever default is live.

The Builder. Monday morning, this is a rebuild. Prompts tuned for Claude's quirks get rewritten for OpenAI's. System prompts, context packing, retry logic, all of it. Months of Claude-specific work becomes sunk cost that the post-hoc justification will quietly ignore. And model swaps at this layer break specific user groups that never showed up in the benchmark suite: non-English speakers, unusual frameworks, odd project setups. Those regressions arrive silently and get reported as "the agent got worse" weeks later. Whoever runs Replit's on-call is going to have a bad month even if the aggregate numbers look better.

Where the council splits. The Researcher takes Masad's benchmark claim at something close to face value, a real production test saying OpenAI is better at this specific task. The Skeptic and the Compute Pragmatist think the benchmark is cover for a pricing decision, that "performance" is third on the list for a reason. That's the real fork. If this is capability, it's bad news for Anthropic's standing in coding. If it's price, it's bad news for every lab's margin, because it means the whales now set the terms.

What it hinges on. One question: did OpenAI win on the eval or on the invoice? We can't see the negotiated rates, so we judge by what happens next. If OpenAI is buying logos at a loss to win reference deployments, the pattern repeats, and soon. Another large agentic coding platform flips its default to OpenAI on the same cost-plus-performance language within two quarters. If this were purely Claude being outclassed at coding, you'd expect the switches to spread across model families and for Anthropic to lose coding-specific accounts on capability grounds, not for OpenAI specifically to keep landing the big token-burners.

Prediction: Before the end of Q1 2027 earnings season in February 2027, at least one more AI coding platform of comparable scale to Replit (Cursor, Lovable, Bolt, or Vercel's v0) will publicly switch its default model to an OpenAI model, citing cost and performance.

Confidence: Medium. The pricing mechanism is real, but timing and who moves first are uncertain.

Why: Replit's switch reads as OpenAI paying up to win a visible, token-heavy reference customer, because agentic coding loops burn output tokens and premium pricing punishes exactly that usage pattern. A lab chasing coding-market credibility will make the same offer to the next big logo, and those platforms run the same economics Replit does, so they have the same incentive to take the discount. The opposite outcome, Anthropic holding every other major coding default, would require OpenAI to stop buying logos right after landing its biggest one, which cuts against a lab that just proved the play works on stage at Disrupt.

Revisit by 2027-02-28: We're right if a Replit-scale AI coding platform publicly makes an OpenAI model its default and names cost or performance as a reason. We're wrong if no comparable platform switches its default to OpenAI by then, or if the only default moves go the other way, toward Anthropic.