Replit makes OpenAI its default and CEO says he is done with Anthropic, but the record is murkier

Did Replit really drop Anthropic for OpenAI? The short version: Replit has made OpenAI its default model for Replit Agent, its CEO said on stage that it is no longer working with Anthropic, and the paper trail suggests something less final than a breakup.

Replit CEO Amjad Masad made the remarks at TechCrunch Disrupt, reported October 17, 2026. By TechCrunch's account, he said the decision was "based on both cost, partnership, and performance," and that "I think ChatGPT is better now." Afterward he clarified that customers can still choose Anthropic models, but OpenAI is now the default.

What changed, and what did not

The default matters more than the rhetoric. TechCrunch notes that most customers keep the default, so Replit's spending with Anthropic could fall significantly. Masad said Replit had recently been a top-five Anthropic customer. It announced its OpenAI partnership in August.

Replit's documentation, last updated October 2, 2026, lists OpenAI, Anthropic, Google, Grok and OpenRouter as providers, with managed usage billed at the provider's public API price. That page does not say which model is the default, and nothing in it removes Claude.

An unnamed person told TechCrunch the Anthropic partnership remains in place. Replit's late-September blog post cited Claude Fable 5 and 5.1 running in Replit Agent's production environment, and a Replit product employee promoted Opus 5.5 on X around the same time. Masad also said "Replit remains model agnostic."

Our reading: a default switch is real. "No longer working with Anthropic" may mean the commercial relationship has ended or shrunk, and the companies have not confirmed either way.

What Replit's own data shows

The evidence for the performance claim is Replit's September 29 post, "Free the models: Harness design at the frontier." It reports results from Replit's own harness, so it is a vendor test, not an independent one.

On DeepSWE v1.1, a 113-task set, Replit Agent scored 72% at $2.11 per task. GPT-6 Astra alone scored 67% at $1.60 at low effort and 74% at $4.43 at the highest. On Terminal-Bench 4.0, with 63 tasks and GPU tasks excluded, Replit Agent scored 49% at $2.53, while Astra scored 42% at $2.25 on low effort and 60% at $5.86 on the highest. Astra's best settings beat Replit Agent by 2 and 11 points at more than twice the cost.

The same post describes delegation behavior. In Replit's production data at medium effort, GPT-6 Astra handed work to a general worker on 20% of turns, against 0.9% for Fable 5 and 2.3% for Fable 5.1. Replit says Astra "is the first model we've seen routinely delegate to general workers without being told to."

Notice what is not in the post: a head-to-head score for the Anthropic models on the same two benchmarks. The tables compare Astra with Replit's own agent. The claim that OpenAI beats Anthropic comes from Masad's remarks and Replit's internal and public benchmarks as he described them, not from numbers we could check.

Cost is the other half of the story

"Cost" came first in Masad's list. Replit passes provider prices through, so a model that scores two points higher at twice the price is not an obvious win for a platform that bills usage. Our API price tracker shows how widely per-token prices differ across the labs, though Replit's cost per task depends on how many tokens its harness spends.

A default is also a commercial lever. When a platform with Replit's reach changes it, one lab gains volume and the other loses a showcase customer. Masad said enterprise was about 2% of Replit's revenue last year and about 15% this year, and that more than half of revenue is business-to-business, so reliability and partnership terms weigh more than a leaderboard.

What to take from it

Three things are established: the default changed, the CEO said what he said, and Replit's benchmarks show Astra performing well in its harness at a price. Three things are not: that Anthropic is gone, that OpenAI's lead holds on independent tests, and that the choice was about quality more than price.

Developers building on Replit should check which model their Agent runs and whether they have pinned a choice. The docs say users can switch provider or bring their own API key, with billing then direct from the provider.

Why platforms switch

For a platform like Replit, the model is an input cost and a quality lever at the same time. OpenAI's August partnership with Replit gave Replit a second supplier and a negotiating position.

That context matters for reading the quote. A CEO announcing a new default on a conference stage is also telling his other supplier that the account is up for grabs. Both things can be true: the benchmarks can favor the new partner and the price can be a bargaining chip.

Our take

Treat the on-stage quote as marketing for the new partner and the benchmark post as a useful but self-selected test. The honest summary is "OpenAI is the default, Claude is still selectable, and the relationship is disputed."

The wider lesson is that coding platforms are the new swing voters between labs. They test constantly and switch on cost, so a single quarter's default can tell you more about price and harness fit than about general intelligence. We will update this piece if either company confirms the status of the partnership.