djcromanDAILY AI NEWS
News

Replit makes OpenAI its default and moves Claude to the menu

Replit CEO Amjad Masad says OpenAI is now the preferred partner and default model for Replit Agent, citing cost, partnership and performance. Claude stays selectable, but most users never touch the default, and Replit was until recently one of Anthropic's five biggest customers.

By djcroman

djcroman news card: Replit makes OpenAI its default and moves Claude to the menu

In the AI coding market, the most valuable piece of real estate is not a benchmark leaderboard. It is the default setting. This week Replit, one of the best known platforms for building apps with AI, moved that setting from Anthropic to OpenAI. CEO and co founder Amjad Masad said on stage at TechCrunch Disrupt that OpenAI is now the preferred partner behind Replit Agent, and that the decision came down to cost, partnership and performance.

The first version of the story sounded like a full breakup. According to TechCrunch, Masad said on Thursday, 15 October, that Replit was no longer working with Anthropic. Two days later he clarified that Claude models remain available on the platform. The truth sits in between, and it says a lot about how money flows between AI labs and the apps built on top of them.

Timeline from OpenAI's Luna price cut on 30 July to Replit's Free Mode in August and Masad's statements at TechCrunch Disrupt on 15 and 17 October 2026

What Masad said on stage

The exchange started with a simple question. The interviewer asked whether the OpenAI partnership Replit announced in August meant the company had stopped working with Anthropic. Masad answered "I mean, yeah," according to Unite.AI's account of the session and IBTimes.

He then explained why. "I think ChatGPT is better now," Masad said, as quoted by TechCrunch. Replit runs an internal benchmark to compare models on its own workloads and also publishes a public one. "We publish a public benchmark as well, and consistently, Sol and Astra perform better on building vibe-coded applications from scratch," he said. "It was a decision based on both cost, partnership, and performance. And for all these reasons, we just felt like they were a better partner for now."

The most striking detail came almost in passing. Masad noted that Replit had been, until recently, one of Anthropic's top five customers. That makes this more than a product tweak. A company that sat near the very top of Anthropic's customer list has changed which model handles most of its traffic.

The clarification: model agnostic, but with a favourite

The reaction was quick. TechCrunch reported that people familiar with the partnership rejected Masad's onstage comments, and one person said the partnership between Anthropic and Replit was still in place. Masad then softened the message. "Replit remains model agnostic," he told TechCrunch. "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."

On Saturday, 17 October, he repeated the point on X: "To clarify: Replit remains model agnostic. Our users can choose Anthropic models as well as many other supported models. What I meant at TC Disrupt is that OpenAI is now our preferred partner powering Agent based on performance, cost, and strength of our partnership." In a reply, quoted by WebProNews, he added that "OpenAI models are generally more affordable."

Replit's own product pages back up the softer version. Its October 2 changelog added new models from both companies, GPT-6.1 Sol in Max Mode and Claude Sonnet 5.5 in Power Mode. And one day after Masad's comments, the October 16 changelog announced that builders on the Core and Pro plans can now pick Claude Haiku 5.5 from the model selector in Power Mode. Claude is clearly still on the menu.

Two columns comparing what changed at Replit, such as the default model and preferred partner, with what stays, such as Claude models in Power and Max Mode

Why the default is the whole game

If Claude is still available, why does this matter? Because almost nobody changes the default. TechCrunch makes the point directly: the shift could significantly reduce Replit's spend with Anthropic, since most customers stick with the default rather than selecting a model themselves.

Replit's pricing design makes this even stronger. In August the company launched Free Mode, and according to its launch post that mode is powered by OpenAI's GPT-5.6 Luna. Core subscribers, who pay 20 dollars a month, can create up to 30 times more than before and get up to 30 hours of chat per month. Everyday tasks in Free Mode no longer use credits, and the allowance resets every five hours.

Fortune reported at the time that Free Mode is the default for Core and Pro users, and that GPT-5.6 Luna is the only model powering it. Replit president Michele Catasta told Fortune the feature was made possible by OpenAI cutting the cost of Luna by 80 percent on 30 July. When a task needs more power, the agent can route it to a stronger model and then fall back to Free Mode. In its own post, OpenAI said harder tasks can go to GPT-5.6 Sol, and noted that Replit was an early user of GPT-3.

Replit's documentation adds one important line: in Free Mode, Replit always picks the model, and users cannot select one. To choose a specific model, including any Claude model, you need to switch to Power or Max Mode, which use monthly credits. So the path of least resistance on Replit leads to OpenAI, and the paths that lead to Claude cost credits.

What Replit's own data shows about the models

Replit is unusually open about how models behave inside its agent. On 29 September its engineering team published a long post titled Free the models: Harness design at the frontier. It describes a harness where the main model, the core loop, decides on its own when to hand work to smaller and cheaper subagents, how hard they should think, and whether to return to a subagent it already briefed.

The post includes one week of production data for three models, each at medium reasoning effort: Anthropic's Claude Fable 5 in August, and Claude Fable 5.1 and OpenAI's GPT-6 Astra in September. Fable 5 dispatched a subagent on 32 percent of turns, Fable 5.1 on 21 percent and Astra on 36 percent. The bigger gap is in delegation of actual implementation work. Astra handed work to a general worker on 20 percent of turns, while the Fable models did so on less than 3 percent. Replit wrote that Astra is the first model it has seen routinely delegate to general workers without being told to.

Bar chart from Replit production data: Claude Fable 5, Claude Fable 5.1 and GPT-6 Astra compared on subagent dispatch, delegation to general workers and reuse of existing subagents

That behaviour matters for cost. Frontier models are expensive per token. A model that keeps the hard decisions for itself and passes routine edits to cheaper subagents can deliver similar quality for less money. In Max Mode with Astra as the core loop, Replit reported 72 percent on the DeepSWE v1.1 benchmark at 2.11 dollars per task, and 49 percent on Terminal Bench 4.0 at 2.53 dollars per task.

The same post is also kind to Anthropic. Replit's engineers wrote that "the code quality of Astra and Fable 5.1 also let us use our code-review subagent less, with no drop in our eval scores." In other words, the quality argument is close. The decision seems to rest at least as much on price and on the shape of the partnership as on raw capability.

A caveat applies to all of this. These are Replit's numbers, measured on Replit's workloads with Replit's harness. As IBTimes notes, they do not prove that OpenAI's models beat Claude on every coding task, and Replit has not published figures showing exactly how much cheaper its preferred models are.

Replit is pivoting to enterprise

Masad also shared where Replit's money now comes from. "We're not sharing ARR numbers just yet," he said on stage, but enterprise contracts went from about 2 percent of revenue last year to about 15 percent this year, and more than 50 percent of revenue is now business to business. Backstage, he told TechCrunch he expects more than half of Replit's revenue to come from enterprise and government customers in 2027.

That shift helps explain the model decision. Enterprise customers care about predictable cost per task, and a coding platform that sells to companies needs margins that do not collapse every time a user asks the agent to rebuild a whole app. Masad framed safety in similar business terms. "Agents need a habitat for them to exist and do good work," he said, describing orchestration, error recovery, sandboxing, networking, identity and authorization as the conditions for deploying AI safely.

Bar chart of Replit revenue mix: enterprise contracts about 2 percent in 2025 and about 15 percent in 2026, B2B more than 50 percent, and a goal of more than half from enterprise and government in 2027

What it means for Anthropic

For Anthropic, the timing is awkward. On 14 October the company told prospective investors that third quarter revenue rose 45 percent to 16.8 billion dollars, as we reported earlier this week. According to Reuters, some investors were disappointed by the slower sequential growth after revenue more than doubled in the second quarter, and some discussed an IPO valuation of around 1.5 trillion dollars instead of the roughly 2 trillion dollars expected before. When the interviewer at Disrupt asked whether losing Replit played a role in that slowdown, Masad only repeated that Replit had been a top five customer.

Nobody outside the two companies knows how large Replit's spend with Anthropic was or how fast it is falling. But coding has been Anthropic's strongest business, and the company's own success with tools like Claude Code has made it both a supplier and a competitor to platforms like Replit. OpenAI, on the other hand, gets a high volume default slot on one of the best known AI app builders, and a reference customer that publicly calls its models cheaper and better for app building.

Our take

This is how the AI price war plays out in practice. It rarely looks like a clean breakup. It looks like a changed dropdown, a new free tier that runs on one lab's cheapest model, and a CEO who first says too much and then clarifies. Replit is right to stay model agnostic, because the lead in coding models can change quickly.

For developers on Replit, nothing breaks. If you prefer Claude, you can still pick it in Power or Max Mode, and you will pay with credits as before. If you never touch the setting, you are now working with OpenAI. For the labs, the lesson is harder: being the best model is not enough if someone else is cheaper by default.

Sources

Source: techcrunch.com

Ai-codingAnthropicOpenaiReplit

Newsletter

The AI news that matters, in your inbox.