Replit’s chief executive said on stage this week that OpenAI has replaced Anthropic as the default model behind Replit Agent. Amjad Masad made the remark at TechCrunch Disrupt on Thursday, and the claim spread across outlets over the weekend. The switch is disputed, and Replit has since softened the wording it used on stage.

The Replit default model change is practical as well as commercial. Replit is a browser-based development platform where people build and deploy applications, so a change to its default can change what gets built. It also gives the fight between OpenAI and Anthropic over coding customers a concrete data point.

What Masad said on stage

According to IBTimes Singapore, Masad told the audience that Replit was “no longer working with Anthropic.” Asked whether the decision followed the OpenAI partnership Replit announced in August, he replied, “I mean, yeah.”

He gave two reasons. The first was performance. Masad said OpenAI’s models currently perform better at building applications from scratch, particularly in the workflows associated with so-called vibe coding. The second was money. He said Replit had until recently been a top-five Anthropic customer, so moving the default away from Claude could cut its Anthropic spending considerably.

Masad summed up the performance argument in a line reported by Innovation Village: “I think ChatGPT is better now.”

The performance case rests on Replit’s own internal and public benchmarks. Reporters did not independently verify those tests, and no dollar figures for the Anthropic spending were published. The cost argument cannot be checked yet.

Replit softened the wording after the event

The strongest version of the claim did not last through the day. In a statement to TechCrunch, Masad said, “Replit remains model agnostic,” and added, “OpenAI is now our preferred partner to power Replit Agent, based on performance, economics, and the strength of our partnership.” Hortolandia News carried the statement in its report.

That shift matters. “No longer working with Anthropic” describes an exit. “Model agnostic” and “preferred partner” describe a default setting. Under the revised account, customers can still pick Anthropic’s Claude or other supported models, and OpenAI is the model that runs when nobody chooses another.

Sources dispute that the Anthropic deal has ended

Many headlines framed the move as a break, including one that read “Replit is no longer working with Anthropic, CEO says”. Sources familiar with the partnership disagreed. They told reporters the Anthropic deal remains in place, and they pointed to Replit’s recent promotion of Claude models.

The record does not settle the question. Reporting says a late-September Replit post named Claude models in production, which fits the idea that Claude still has a role inside the product. Neither Anthropic nor OpenAI had given an on-record response in the reporting reviewed. Until one of the companies publishes specifics, the accurate description is that Replit changed its default and has not confirmed a full exit.

What the switch means for people building on Replit

A default is a starting point. The change affects which model Replit Agent uses unless someone picks another, and most customers never change defaults. Defaults shape outcomes for everyone who does not check, which is why a quiet setting can move a lot of work without a formal announcement.

Reporting names two OpenAI models, Sol and Astra, as the ones Masad had in mind. For context, a Microcenter weekly AI roundup dated October 2 covered OpenAI’s GPT-6.1 Sol launch, about two weeks before Masad’s remarks. Reporting on the switch, including a dev.to write-up, notes that Claude remains available.

Developers who rely on specific Claude behavior may notice differences in the code Agent produces. Those differences can affect test results, dependency choices and edge-case handling, so anything that shipped on Claude deserves a second look.

A separate analysis from beri.net argues that Replit’s standard data agreement does not appear to require advance notice of model changes. For teams that treat the model as part of their compliance review, that gap matters more than which model is the default.

Why the OpenAI and Anthropic rivalry matters for coding customers

If the switch holds, it gives OpenAI a visible win in coding and takes a top-five customer away from Anthropic. Investors and procurement teams are likely to read it as a signal about coding-tool market share, though one platform does not settle a market. The move also arrives during a broader debate over coding-agent economics and model pricing among the major labs.

Regulators are watching the labs as well. The FTC has sent civil investigative demands to OpenAI and Anthropic, a separate matter from the Replit remarks. Coding agents also touch sensitive data. India’s CERT-In has said AI agents can read live business records without anyone noticing, which is a reminder that admins need to know what is running.

What business customers should check this week

Business customers have the most leverage here, and the steps are concrete:

  • Open the admin console and review which models your workspaces can reach. Restrict access to the models your team has tested.
  • Re-test anything built on Claude before trusting it on the new default. Run the same prompts, tests and security checks against both models.
  • Log which model produced each build so regressions can be traced quickly.
  • Ask Replit in writing whether it will give notice before future default changes, and which Claude models remain in production.
  • Review your data agreement and subprocessor terms, since the analysis suggests advance notice is not required.

Individual developers should choose their model deliberately if Claude works better for their projects, because a default will not keep a personal preference in place. Agent permissions are becoming a governance question in other sectors too. Big banks have already moved to block AI agents from moving customer money.

What to watch next

Three things will settle the story. First, whether Replit publishes usage or spending figures that back its cost argument. Second, whether Anthropic or OpenAI comments on the record. Third, whether Replit names the Claude models it still runs in production.

If the answers show a full exit, Anthropic loses a notable coding customer and OpenAI gains a clear win. If they show a default change only, the story is less about one contract and more about who controls the starting point for everyone who never changes a setting. Replit’s next changelog and any pricing update will say more than the stage remarks did.