Skip to content

Developer tools

Model routing becomes the default across Replit accounts

5 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: Coding agents →
Image: thenewstack.io

Replit has made its intelligent model routing the default on every account. Each session now starts in Free Mode, where a system called Auto chooses the model for each Agent task and can switch to a different one partway through. Replit says the router weighs quality, speed and cost on each decision.

Core and Pro subscribers can still override it and pick models themselves, but only by moving to the modes that cost money. Michele Catasta, Replit’s president and head of AI, described the economics behind the change to The New Stack on August 27: “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.”

How Auto picks a model

Free Mode always runs on Auto, and while it is active you cannot choose a specific model. Most of the work goes to OpenAI’s GPT-5.6 Luna, the smallest and cheapest tier of the GPT-5.6 family, which sits below Terra and the flagship Sol. OpenAI says that when a task needs more advanced reasoning, Replit can hand it to GPT-5.6 Sol and then return to Luna without losing project context. That is the mid-task switching Replit describes: the system revisits its choice as an Agent task develops, looking for a better balance between capability and cost at each stage.

A second layer of escalation reaches past the free models. According to Replit’s documentation, when a paid mode would handle a request better, Agent says so and gives you two options: allow the paid mode for that one request, or stay in Free Mode. A one-time escalation does not stick. Once the request is done, the session goes back to Free Mode.

Catasta says the routing problem comes down to inference. Before choosing a model, the router has to work out the nature, difficulty, scope and intent of a request. Replit trains its routers on proprietary usage data and watches those signals across its user base. It also controls the layer the router sits in. “Replit has owned, from the start, both the agent harness and the infrastructure surrounding models which in turn allows us to train sophisticated model routers,” Catasta said. The harness is the agent’s scaffolding of tools, prompts and orchestration around the model.

He also says the router has been in beta for a long time, alongside experiments with early versions of Auto and with subagent routing. “The most important learning is understanding from first principles the failure modes of every experiment, so we could keep hill climbing on the final system that we just shipped,” he said.

Free, Power and Max

Free Mode launched last week as a lower-cost Agent mode that uses no usage credits. It comes with limits, though. On Core and Pro the allowance resets every five hours and Pro gets more of it, and Settings → Usage shows what is left and when it refreshes. Starter keeps its existing capped access. The “free” in Free Mode is a quota rather than a price, and Replit’s own comparison is that Core subscribers can create up to 30 times more than before and get up to 30 hours of chat a month in it. The company has not published the workload, prompt mix or methodology behind those figures.

Above Free Mode are two paid tiers. Power, formerly called Economy Mode, balances cost against capability. Max uses higher-performance models for larger changes, deeper reasoning and longer builds. In either one, Core and Pro users can turn Auto off and choose a model, or leave Auto on and let Replit route within the paid tier. Replit Routines, which automate recurring tasks, do not run in Free Mode at all. They need Power or Max and a budget for each run.

Paid Agent work is priced by effort rather than by a token meter per task. The charge covers model calls, tools, orchestration and the finished work, and paid actions ask for confirmation before they start. Deployment, storage, databases and other cloud services are billed separately.

Enterprise accounts get a different control: administrators can limit Auto to an approved set of models for each workspace. Routing still happens automatically, but only within the models company policy allows.

A crowded month for routers

Replit is not alone. Earlier in August, Stripe agreed to acquire the model gateway platform OpenRouter for a reported $8 billion, and Ramp launched Router.com the same day. Router.com sends requests to the cheapest model that meets a performance bar you set. In July, Cursor shipped a router that picks models for coding requests and claims comparable performance at substantially lower cost. Meta is reportedly building an internal router called “Switchboard” that scores coding tasks by difficulty and sends the simpler ones to cheaper models.

Cursor, which SpaceX agreed to buy for $60 billion, has spent years building its own coding models, including the Composer family, and has kept working on more advanced ones under SpaceX. Replit is not building its pitch around owning models. It is betting that control of the agent and the systems around it gives it enough information to choose models well as tasks run.

Catasta calls routing an early-stage field and only one part of a larger research effort. “Routing is but one piece of the puzzle that is tightly integrated to many other aspects of our harness research,” he said. He is less modest about what Replit’s usage data adds compared with a standalone gateway: “No third-party router company could reproduce the same results for our own agent.”

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.