Replit's Move to OpenAI Is Now Confirmed as a Deal Worth More Than $100 Million a Year, Sealed With a Discount Above 30 Percent

The mysterious part of the Replit-Anthropic split is now the priced part. The Information reported on October 24, citing a source, that Replit recently moved most of its AI spending from Anthropic to OpenAI in an arrangement worth more than $100 million a year, and that OpenAI's discount of more than 30 percent closed the deal after Anthropic discontinued the discount it had been giving the coding platform.

The report turns a murky episode into a confirmed one. Replit CEO Amjad Masad said on October 17 that his company had stopped working with Anthropic, citing cost, partnership quality and benchmark wins for OpenAI's Sol and Astra models on apps built from scratch. A person familiar with the relationship disputed the split at the time, and Replit's own September engineering post showed Claude models still running in production that month. What is now clear is that the switch was real and had a number attached: a nine-figure annual commitment, decided by a 30-percent-plus gap in pricing.

The same reporting describes a broader pattern that should worry Anthropic's enterprise sales team more than any single defection. Parker Conrad, CEO of HR software maker Rippling, said that in May, Anthropic models handled 57 percent of the coding work across his company's hundreds of engineers while OpenAI handled 16 percent; the balance has since reversed, with 52 percent of coding tasks now on OpenAI models and 18 percent on Anthropic, a shift Conrad attributed to cost. DataCamp, a training provider with roughly 200 employees and a heavy Anthropic buyer, received a 40 percent discount from OpenAI, according to the report, on the condition that it adopt Codex rather than rival tools like OpenCode or Claude Code.

Discounts of this size were rare when frontier-model capacity was scarce, and their appearance now says something about how the market has changed. Application-layer products like Replit wrap whatever model they call behind harnesses, retries and scaffolding, which makes the models themselves more substitutable than benchmarks imply. When two frontier options are close enough in quality, a 30 percent difference in price becomes the whole decision, and a seller nearing an IPO has less room to let nine-figure accounts walk away over discount policy.

For Anthropic, the timing is awkward. The company is reportedly weeks from a public listing, and its coding franchise has been the loudest proof point in its enterprise pitch. Losing share in exactly that category, to a competitor openly buying share with discounts and bundling conditions, gives short-sellers a ready-made narrative. For OpenAI, the wins are a counterweight to months of stories about rival labs taking its enterprise customers, and evidence that its pricing power strategy is producing measurable results.

The caveat is that the reporting rests on anonymous sources and company executives speaking for their own procurement decisions, not on confirmation from Anthropic or OpenAI. But if the numbers hold, the most important benchmark in the model business this quarter is not a leaderboard. It is the invoice.