Replit reportedly moves most AI spending from Anthropic to OpenAI

What happened

According to The Information, Replit moved most of its AI spending from Anthropic to OpenAI. The report, dated 24 October 2026, describes a contract worth more than $100 million a year. It says OpenAI offered a discount above 30 percent after Anthropic would not extend a discount Replit had received earlier unless Replit agreed to spend more. Neither company has publicly confirmed these figures, and this article relies on the report's account.

The switch became public a week earlier. On 16 October 2026, Replit's chief executive and co-founder, Amjad Masad, told an audience at TechCrunch Disrupt in San Francisco that the company was no longer working with Anthropic. He gave cost, partnership and performance as the reasons. In the days that followed, he clarified that customers could still choose Anthropic's Claude models, while OpenAI's ChatGPT models had become the default for Replit Agent. TechCrunch reported that Masad told it Replit remains model agnostic.

The picture is not fully settled. A person familiar with the arrangement told TechCrunch that the Anthropic partnership is still in place, and Replit's documentation still lists several Claude models. Read together, the reports point to a change in the default model and in where most of the spending goes, not to a clean exit from Anthropic.

Why it matters for businesses that automate work

Many automated workflows, from drafting follow-ups to sorting expense claims, call an AI model in the background. The owner of the workflow may never have picked that model. The supplier chooses it, and the supplier can change it when its prices or contracts change. Replit's case shows how a default setting can move most of a platform's usage to a different provider, because most users keep whatever the default is.

The same report also points to a pattern in pricing. The Information's account of Rippling, an HR software firm, shows how quickly a coding workload can move. In May 2026, Anthropic models handled most of the coding work done by Rippling's hundreds of engineers. The shares have since reversed.

Rippling's CEO, Parker Conrad, attributed the change to cost, according to the report. Smaller buyers have also reported discounts. TechCrunch described a training company, DataCamp, that received a 40 percent discount from OpenAI on the condition that it adopt Codex.

What changes in practice

For most owners, this does not mean changing software this week. It means knowing what sits underneath the tools you already use. These steps apply whether you run five people or five thousand.

1. List every tool that calls an AI model, including tools bought as part of a larger package.
2. Ask each vendor which model is the default, when it last changed, and whether it can be fixed in your account settings.
3. Write your prompts, approval rules and expected outputs in your own documents, so that a second model can be tested against the same standard.
4. Track AI charges by provider each month, so a shift in spend is visible on your own books.
5. Before a model change goes live, run it on a sample of real work, such as invoices or customer messages, and compare the results with your current output.

Where a contract is large enough, the price terms can be negotiated. Most businesses will not have that leverage. The practical protection is a record of what runs where and a test process that you control.

What automation cannot do here

Automation can keep a record of which model runs each task and can flag when a default changes. It cannot judge whether a new model is accurate enough for your invoices or your customer replies. A person still needs to test outputs and approve any switch.

What to watch next

• Whether Replit or OpenAI publicly confirms the contract terms, or whether they stay unconfirmed.
• Whether Anthropic responds with its own pricing for the accounts named in the report.
• Whether the defaults in Replit's documentation change again, and whether Claude remains selectable.
• Whether other software companies report similar shifts in their own model usage.