Send a prompt in the shipped configuration and you do not name the model that answers it. Every account starts in Free Mode [4], and Free Mode offers no manual model selection [7]. Auto reads the task, weighs quality, speed and cost, and can revise its pick as the task evolves [2]. To choose yourself you need a Core or Pro subscription, and you have to switch into Power or Max and turn Auto off, which can incur usage costs [8]. Free Mode consumes no usage credits and its limits reset every five hours, with more headroom for Pro [6]. Free of credits, not of limits. Replit may also nudge a task into one of the paid modes when it decides more capability is needed [9]. So the override is real, and it lives on the metered side of the meter [20].
The stated reason is arithmetic, not capability. Michele Catasta, Replit's president and head of AI, told The New Stack that per-token rates inside a single model family can span orders of magnitude, and that cheaper small models have moved much closer to frontier ones [15]. Both halves have to hold on your workload before the saving reaches your invoice. Your task mix has to be dominated by work a small model finishes on the first attempt, and the difficulty estimate has to be accurate enough that the retries it causes cost less than the tokens it saves. Neither figure is published. Cursor's router, launched in July, claims comparable performance at substantially lower cost [13]; that is a claim about Cursor's request mix, not about yours.
The build discipline reads well. Catasta says the system ran in beta for a long period, and that the most important learning was understanding the failure modes of every experiment from first principles [16], after months of earlier work on Auto mode and subagent routing [17]. Enumerate the misroutes, then optimise. That is the correct order.
The market has now priced this layer twice in a month. Stripe agreed to pay a reported $8 billion for the model gateway OpenRouter [11]. SpaceX agreed to pay $60 billion for Cursor, which has long invested in its own coding models, including the Composer family [18]. The gateway that only chooses models fetched roughly an eighth of the company that also trains them [21]. Replit is explicit that owning the model layer is not its pitch; it is betting that controlling the agent and the systems around it is enough [19]. Ramp's Router.com routes to the lowest-cost model that clears a specified performance bar [12], and Meta is reportedly building an internal router, Switchboard, that scores coding tasks by difficulty and sends the simple ones to cheaper models [14].
For anyone maintaining prompts, the unit under test has changed. You used to evaluate a model sitting behind an agent. In the default mode you are evaluating a routing policy that can hand the same task to a different model as it progresses [2]. If your regression suite assumes a fixed backend, that assumption is now a setting someone else owns. The honest way to run this is to keep one pinned-model lane on Power or Max for the evals you trust, accept the credit cost of that lane, and let Auto have everything else.
