Model routing becomes the default across Replit accounts

Replit has made its intelligent model routing system the default on every account: a feature called Auto mode now decides which underlying model handles each task, weighing quality, speed and cost as work proceeds. Every user starts in Free Mode with the router choosing on their behalf, according to The New Stack.

At a glance

• The router evaluates quality, speed and cost, and can switch models partway through an Agent task when the trade-off between capability and expense shifts as the work progresses.
• Core and Pro subscribers keep an escape hatch: Power and Max modes let them turn Auto off and pick a model, though those modes can consume usage credits.
• Enterprise administrators can restrict Auto to an approved set of models for each workspace, so automatic routing continues while model choice stays inside whatever policy the organization has set.

Routing is becoming the layer where AI coding vendors compete on unit economics rather than raw capability. Making the router the default reads as a bet that most agent work does not require frontier models, and that the margin sits in matching task difficulty to price. For developers it also means less visibility into which model produced which piece of work, unless they move up to the modes that restore manual selection.

Free Mode resets its usage limits every five hours

The groundwork surfaced last week, when Replit introduced Free Mode, a lower-cost Agent mode that does not consume usage credits and leaves model selection to Auto. It is not unlimited: Replit offered it to Core and Pro subscribers with limits that reset every five hours, and larger allowances for Pro.

Free Mode users cannot select a model by hand. Core and Pro subscribers who want that control switch to Power or Max, where Auto can be turned off, and Replit may itself suggest moving a task into one of those modes when it determines that more capability is required.

Michele Catasta, Replit's president and head of AI, said the company spent recent months experimenting with early versions of Auto mode, subagent routing and multiple iterations, and tested intelligent model routing in beta for a long period before releasing it publicly. The key learning, he said, was understanding the failure modes of every experiment from first principles.

Stripe agreed to acquire OpenRouter for a reported $8 billion earlier in August

The change lands amid a rush of routing activity. Earlier in August, Stripe agreed to acquire model gateway platform OpenRouter for a reported $8 billion, and Ramp launched Router.com the same day, sending requests to the lowest-cost model that clears a specified performance bar.

In July, Cursor launched its own router, which selects models for coding requests and claims comparable performance at substantially lower cost. Cursor, which SpaceX agreed to acquire for $60 billion, had already invested for years in its own coding models, including its Composer family.

Under SpaceX, Cursor has been developing more cutting-edge models as well. Meta, meanwhile, is reportedly developing an internal router called Switchboard that scores coding tasks by difficulty and sends the simpler jobs to cheaper models. Replit is not putting ownership of the underlying model layer at the center of its pitch.

Replit owns the agent harness and trains routers on its own usage data

Catasta told The New Stack that the wider push into routing comes down to economics, and that Replit has owned both the agent harness and the infrastructure surrounding models from the start, which in turn lets it train sophisticated model routers.

Across one model family, per-token rates can span orders of magnitude. At the same time, the intelligence of cheaper, smaller models is now much closer to their larger frontier counterparts, providing us a lot of room for cost optimizations.

Catasta said a router must infer the nature, difficulty, scope and intent of a request, and that Replit trains against proprietary usage data and observes those signals across its user base. No third-party router company, he said, could reproduce the same results for Replit's own agent.

What Catasta says routing still lacks

Catasta called model routing an early-stage field and said further research is expected to move the needle on serving the best intelligence when customers need it, describing routing as one piece of a puzzle tied to Replit's broader harness research. Replit has not published figures for what Power and Max cost in usage credits, nor the exact allowances behind the five-hour Free Mode reset.