Replit Prefers OpenAI: What Users Can Control When Choosing Models

Replit now prefers OpenAI for its coding agent while continuing to offer models from other providers. CEO Amjad Masad clarified that position on October 17. For users, the practical question is what automatic selection handles for them, and when choosing a model themselves is worthwhile.

Key takeaways

• Masad says OpenAI is Replit’s preferred partner; that does not amount to removing Claude entirely.
• Free Mode selects models automatically. Eligible users can choose manually in Power and Max.
• Available options depend on the account and workspace policies.
• A comparison becomes more meaningful when the task, starting state, and success criteria stay the same.

The default is a product decision in its own right

Masad attributes the preferred partnership to performance, cost, and the relationship between the companies. His clarification is readable as an embedded original statement on Digg. TechCrunch also reported the change. The established point is therefore the platform operator’s preference, rather than an independent finding about which model wins every coding task.

That distinction is useful for people who want to turn a description into an application. They are using an environment in which tasks are planned, files changed, and results reviewed, rather than buying an isolated text generator. A model can produce a good draft and still require more rework. Conversely, an unremarkable response can be exactly right for a small change. Both are reasons to judge the finished result.

A default initially removes a decision for the user. That can make getting started easier: someone building an internal scheduling overview need not study model rankings before their first attempt. It also means the platform can change its selection while the user continues opening the same product. Our interpretation is that the design of that selection is itself becoming an important part of an AI tool’s quality.

Replit’s August announcement of Free Mode already described closer collaboration with OpenAI. Its promise of doing more with an existing subscription was a product claim, rather than a measurement for an individual project. Likewise, the current preference does not establish a specific saving on the next bill.

How users can influence the selection

Replit’s current documentation distinguishes Free, Power, and Max. Free always uses automatic model selection. Power is aimed at cost-conscious work, while Max targets more demanding tasks. Auto in Power or Max also consumes subscription credits; it should not be equated with Free. That distinction matters more in everyday use than the word “automatic” alone.

Core or Pro users who have access to the feature’s rollout can open Agent Mode in the prompt box, choose Power or Max, and then pick an available option under Primary model. The choice applies to the next message. If Primary model is missing, the feature is not yet available on that account. The documentation lists options including GPT-6 Sol and Claude Sonnet 5 for Power, and GPT-6 Astra and Claude Opus 5.5 for Max.

A specific model name still does not describe the entire assignment. In its September engineering article, Replit explains that the main agent can also delegate work to specialized subagents. The surrounding software, often called the harness, defines its tools and available actions. Assessing results therefore means looking at the entire process. Similarly, Amp’s expansion into complete work assignments shows why the benefit extends beyond writing individual lines of code.

Businesses have another layer to consider. Replit’s documentation says Enterprise administrators can control the available model pool for each workspace. Automatic selection remains restricted to the authorized pool. A missing option may therefore reflect an organizational policy. The general public model list is not a promise that every option will appear in a particular workspace.

A useful comparison starts with a small task

A practical approach is to choose a limited change: for example, adding a search field to an existing list, including an empty result and an understandable error message. Decide beforehand what counts as completion. Without those criteria, the nicest explanation can easily beat the most useful implementation. This is a recommendation for evaluation, rather than a product test we conducted.

Repeating the task should use the same starting state. Replit documents checkpoints and returning to earlier development states for that purpose. It also identifies an important boundary: a normal rollback does not reset databases by default. The development database can be included separately, while the production database is not automatically restored through that feature. A comparison must therefore establish which data should remain unchanged.

It helps to record rework specifically: which requirement remained unmet, which function worked, and which additional step was needed? A short table outside the application is enough. That creates a project-specific decision without requiring a global ranking or a change of provider. Reaching the goal reliably with the default is just as useful an outcome as deliberately switching models for a particular task.

The opportunity ahead is better assistance with selection. Replit’s preferred partnership makes the automatic choice more visible, while continued model choice leaves room for users’ own experience. The biggest advance would be an environment that connects suitable steps with understandable effort. The model brand leading that process becomes interesting when it actually moves the user’s project forward.