Andy Garcia · freelance dev
FRENES
Back to the blog
AI & Platforms

Replit Says OpenAI Is Now Its Default. Seven of the 11 Models in Its Picker Are Still Claude, and Free Mode Lets You Choose None.

October 19, 2026 6 min read

On Thursday, October 15, onstage at TechCrunch Disrupt in San Francisco, Replit CEO and co-founder Amjad Masad was asked whether the partnership Replit had announced with OpenAI meant it was no longer working with Anthropic. "Yeah. I mean, yeah," he answered, in the session video TechCrunch posted on YouTube. He gave three reasons: cost, partnership and performance. He said he thought OpenAI's models were better now, cited an internal benchmark and a public one that Replit publishes, and said Sol and Astra consistently did better at building vibe-coded applications from scratch. TechCrunch reports that he also described Replit as having been, until recently, one of Anthropic's top five customers.

Then came the clarification. After the event, TechCrunch reported, Masad said customers can still use Anthropic's models, but OpenAI is now the default. Over the weekend that followed, he told TechCrunch: "Replit remains model agnostic. Our users can choose Anthropic's Claude or other supported models … OpenAI is now our preferred partner for powering Replit Agent, based on performance, economics, and the strength of our partnership."

The other side is thinner. TechCrunch writes that insiders familiar with the partnership rejected Masad's onstage comments; one, who asked not to be named, said the partnership was still in place. The story, last updated on October 19, carries no on-record statement from Anthropic. And TechCrunch has retitled it: the URL still says Replit "is no longer working with Anthropic", the headline now says Replit "makes OpenAI its default over Anthropic".

That retitle is the real story. Nobody removed Claude. The default moved, and on a platform that picks models for you, the default is the decision. TechCrunch says it plainly: most customers stick with the default rather than selecting a model themselves.

I build a LegalTech product on Claude, so a top customer moving its default away from Anthropic is news about my supplier. It is also news for anyone whose product sits on a platform that chooses the model on their behalf. So I read Replit's blog and documentation instead of the quotes.

The switch was on Replit's blog before it was onstage

TechCrunch notes that Replit was still promoting Anthropic products a few weeks before Disrupt. Replit's own posts show a switch made in steps.

On August 18, Replit launched Free Mode, "powered by OpenAI's GPT-5.6 Luna". The post says Core subscribers, on a plan it prices at $20 per month, can "create 30X more" without spending credits on everyday tasks. On August 26, Replit made Intelligent Model Routing available to everyone: Replit picks the model for each task, and in its own testing the router delivered "the same output quality at 65% lower cost than the previous version of Max Mode."

On September 29, an engineering post titled "Free the models" published a production table: one week per model, Claude Fable 5 in August, Fable 5.1 and GPT-6 Astra in September. In the same post, Replit benchmarked Max Mode, described as "our highest-quality setting with Astra as the core loop". Anyone who read that post could see where Replit's top tier was heading two weeks before Disrupt.

Where the choice really sits

"Customers can still pick Claude" is true, under conditions that Replit's documentation spells out.

Free Mode "always uses intelligent model routing", and builders in Free Mode "cannot use the Model selector or choose a model manually." Manual choice exists only in Power Mode and Max Mode, which spend credits, and the docs offer it to builders on the paid Core and Pro plans. On the Starter plan, you have to upgrade first.

Inside the picker, Claude is still the majority. The model selector reference lists 11 models. Power Mode: GPT-6 Sol, GPT-6 Luna Fast, Claude Sonnet 5, Claude Sonnet 4.6. Max Mode: GPT-6 Astra, Claude Fable 5.1, Claude Opus 5.5, Claude Opus 4.8, Claude Opus 5, Claude Opus 5 Fast, Kimi K3. Seven of 11 are Claude. The docs add that availability "can vary by rollout, organization settings, and authorization."

The Enterprise settings hold the line I found most interesting. Admins approve providers and models per workspace, with OpenAI, Anthropic, Google, DeepSeek, Kimi and Z.ai listed, but "at least one model from each of OpenAI, Anthropic, and Google must remain enabled." A company whose policy allows a single AI vendor cannot express that in those settings. On Replit, "model agnostic" is a product rule, not only a slogan.

That matters because enterprise is where Masad says Replit is growing. Onstage, he put enterprise contracts at about 2% of revenue last year and about 15% this year, with more than 50% of revenue from business customers. He told TechCrunch he expects enterprise and government to bring in more than half of revenue in 2027. Those are the buyers who read model policies.

The benchmark I could check, and the one I could not

Masad cited an internal benchmark and said Replit publishes a public one. The public benchmark Replit names on its own blog is ViBench: in a June 23 engineering post, it calls ViBench "our public benchmark for vibe coding": a plain-English spec, an app built from scratch, then test plans run against the result.

As of October 19, the results page at vibench.ai reads "Last updated May 2026". Its latest snapshot, a Zero-to-One sweep across 24 applications, has no Sol, no Astra and no Fable row. In that snapshot, Claude Opus 4.8 leads GPT-5.5: 87.8% of test plans fully passing at $2.04 per run, against 86.5% at $2.42.

That does not contradict Masad. Those are older models, and newer ones may well flip the order. It means the public comparison behind his claim is not on that results page yet. The benchmark's public code repository did move to a ViBench 2.0 beta on October 6, with Sol, Astra, Fable 5.1 and Opus 5.5 among its builder models, but its README and changelog publish no model scores.

The September 29 post does carry hard numbers, but they answer a different question. On DeepSWE v1.1, Replit Agent with Astra as the core loop scores 72% at $2.11 per task. Astra on its own, at its "xhigh" effort setting, scores 74% at $4.43: two more points for 2.1 times the cost. On Terminal-Bench 4.0, it is 49% at $2.53 against 60% at $5.86: eleven points for 2.3 times the cost. The same post shows the same setup, with a single long-lived worker in place of the subagents, losing 11 and 16 points. Those runs compare harnesses around one model. The post does not publish the same runs with Claude as the core loop.

In an earlier piece, I read the Opus 5.5 and GPT-6 Sol launch pages side by side and concluded that per-token prices say little about cost per task. Replit's numbers make the same point from the platform side. When Masad says "economics", the bill depends on the harness and the routing, not only on a lab's price list.

What I would check before letting a platform pick my model

Four things.

Know your default, in writing. If a platform routes for you, ask which model handles your traffic now, and whether you are told when that changes. Replit says it notifies users before work escalates to a mode that costs credits. The pages I read do not mention a notice when the routed model itself changes.

Pin what you have to defend. Where a choice exists, pin the model for any output a client or a regulator might question, and leave routing to drafts. On Replit, that means Power or Max Mode on a paid plan.

Run your own comparison. Replit's Compare models, in beta and supported only in Chat, runs one prompt on up to four models, each as a separate agent, and the docs warn that "a Compare 4 run costs about as much as four single runs." Five real tasks from your backlog across four models is 20 runs. That is a small price for evidence about your own work rather than a vendor's.

Keep the switch on your side. In the LegalTech I build, which provider processes a client document is a contract question, not a preference. The principle is simple: the model should be a setting I own, and changing it should go through evals and through the client paperwork, never through someone else's announcement. I apply the same habit to writing. The posts I draft with Claude are fact-checked by Jev, a judge model from another company, so I always know which model said what.

Replit did not drop Claude. It changed what happens when nobody chooses, and most customers, TechCrunch notes, leave that choice to the default.

Sources

A project like this one?

I design and deploy products like this. Let's talk.

Let's talk