AI coding company Replit is making its intelligent model routing system the standard setting for every user account, automatically selecting which underlying AI model handles each task as work progresses. The feature, called Auto mode, will become the baseline experience for all accounts, though Core and Pro subscribers retain the ability to manually pick specific models when they want direct control. The system evaluates quality, speed and cost when deciding which model to route tasks to, aiming to match the right level of capability to each job.
The shift follows a wave of similar moves across the AI coding sector. In August, Stripe announced plans to buy model gateway platform OpenRouter for roughly $8 billion, while Ramp unveiled Router.com on the same day to direct requests to the cheapest model meeting performance requirements. A month earlier, SpaceX-owned Cursor rolled out its own router for coding requests, claiming it delivers similar results at far lower expense. Meta is also reportedly building an internal system dubbed "Switchboard" that grades coding tasks by complexity and assigns easier work to less expensive models. Replit says it's been testing early versions of Auto mode and subagent routing for months, with the company learning to identify failure patterns in each experiment before finalizing the system it just released.
Auto mode is now the foundation of Replit's Free Mode, a lower-cost Agent option that doesn't consume usage credits and lets the platform choose the model on the user's behalf, subject to usage caps. Free Mode launched last week for Core and Pro subscribers without drawing down their credits, but imposes limits that reset every five hours, with Pro users receiving higher allowances. In Free Mode, users can't manually select a model. Core and Pro subscribers can switch to Replit's Power or Max modes to disable Auto and pick a model themselves, though those modes can incur usage costs. Replit may also prompt users to move a task into one of those higher-powered modes when it determines more capability is needed. For Enterprise customers, administrators can limit Auto to an approved list of models for each workspace, letting Replit continue routing automatically while staying within company policy.
Michele Catasta, president and head of AI at Replit, tells the publication that the broader industry push into routing stems from basic economics—the widening gap between what models cost and the level of intelligence developers truly need for a given task. "Across one model family, per-token rates can span orders of magnitude," Catasta says. At the same time, cheaper and smaller models now deliver intelligence far closer to their larger frontier peers, creating substantial room for cost savings. Replit argues that its control over both the agent framework and the infrastructure around models lets the company train sophisticated routers that continuously adjust which model to use as an Agent task unfolds, seeking a better balance between capability and expense at each stage. Catasta says a router must infer the nature, difficulty, scope and intent of a request, with Replit able to train against proprietary usage data and track those signals across its user base. "No third-party router company could reproduce the same results for our own agent," he notes. The company contends that model routing remains in early stages, with further research expected to advance how agents deploy intelligence when customers need it most. Routing, Catasta explains, is only one piece of a puzzle tightly woven into many other facets of the company's agent research. Companies chasing efficiency through routing may discover that the real competitive edge lies less in choosing the cheapest model than in understanding the full arc of a developer's intent—a capability that requires deep integration rather than a drop-in gateway.