Amjad Masad was on stage at TechCrunch Disrupt on October 17 when the question of which models sit behind Replit’s coding agent came up, and his answer reversed what the company had been saying for most of the past year.
Replit is no longer working with Anthropic, the chief executive said. OpenAI is the default provider for the platform’s agent, a choice he attributed to cost, to the working relationship between the two companies, and to how the models performed on Replit’s own internal tests. Claude remains available to users who want it.
His summary of the comparison was blunt. ChatGPT, he said, is better now.
The word that matters in that sentence is default. Most users of a coding tool never open the settings, and a default is a routing decision made once by a product team and executed millions of times afterward. When a platform switches its default, it moves volume that no advertisement can buy, and it hands the winning vendor the feedback loop that comes from seeing how a model fails on real work.
The economics explain part of the decision. Agentic coding is expensive to run: a session that plans, edits files, runs tests and retries after failures can consume many times the tokens of a chat conversation, and the platform pays for those tokens before it bills the customer. A price difference of a few dollars per million tokens becomes a margin question at the scale Replit operates.
Cost alone rarely decides these switches, though. Model vendors compete on credits for startups, on engineering support during integration, on early access to new versions and on whether an executive takes the founder’s call at midnight before a launch. Masad listed the relationship alongside price and performance, which is where partnerships usually sit in a decision like this.
Internal benchmarks are the third leg, and the one that is hardest for outsiders to check. Every platform that routes work to models runs its own evaluations on its own traffic, and those tests measure the tasks its customers actually attempt rather than the problems a vendor chose for a public leaderboard. A model that leads a published suite can lose on a codebase with a decade of technical debt.
Anthropic’s position in coding has been strong, which makes the switch worth noting. Claude models have been the default in several coding tools, and the company has built a business around developers who trust it with large repositories. Losing a default at one platform is a data point rather than a verdict, but defaults are how a lead in developer mindshare gets converted into revenue.
The broader pattern is that the model market is separating into layers. At the top, a few labs sell frontier capability at prices that reflect the cost of training. Below that, a widening group of providers sells capability that is good enough for most tasks at a fraction of the price, and platforms route between them based on the job at hand.
Multi-model platforms have learned to keep that routing flexible. A well-built product puts an abstraction layer between its features and any single vendor, tunes prompts per model, and maintains an evaluation harness that can be pointed at a new release within days. The engineering cost of that flexibility is real, and it is what allows a company to change suppliers without changing its product.
For Replit, the calculation involves where its growth is. The company sells a browser-based environment where students, hobbyists and small teams build and host applications, and its agent has become the part of the product that people pay for. Buyers in that segment care about whether the tool finishes the job, not about which lab supplied the model.
There is a risk in being explicit about the choice. Telling users that one model is better invites them to test the claim, and coding tools are easy to compare because the output either runs or does not. The compensation is that a clear default produces consistent behavior, and consistency is what turns a demo into a habit.
Vendors on the losing side of such a switch tend to respond with performance rather than argument. Anthropic has released models on a rapid schedule and has courted enterprises directly, where contracts are longer and switching costs are higher than on a self-serve platform. A consumer-facing default is a smaller prize than an enterprise deployment, and the company has been clear about which it values.
What to watch is whether the switch holds through the next model release. Platforms that route by benchmark tend to revisit the decision when a competitor ships something faster or cheaper, and the ability to switch back is the thing that keeps prices down for everyone.
For Replit’s users, the practical effect this week is a different model answering the same prompt. For the market, it is one more data point that the default is a decision a customer can reverse.
