A model router built for your workload
Train, deploy, monitor, and improve an LLM router. Each router ships as a single endpoint you call like any OpenAI or Anthropic model.
Shift-1, our router for coding agents
We trained Shift-1 on AC2 Routers using our own production coding traffic. It drops into Codex and Claude Code and picks the cheapest model that can handle each request, with no observed drop in intelligence.
$ curl -fsSL https://shift.ac/install.sh | shRoute at every request.Long agentic sessions mix planning, exploration, implementation, and debugging. The router decides at each request, not once per task, so each step gets the model it needs.
Start strong, step down.Starting cheap and escalating when stuck bloats the context and confuses the stronger model. AC2 routers bias toward the strongest model first and step down once the remaining work fits a smaller one.
Switch only when it pays.Changing models mid-session gives up cached input tokens. Routers account for switching costs and only route when it's worth it.
Start anywhere.Begin with a pretrained decision model, fine-tune one on an open-weight base, or bring your own rules.
Routers plug into training and inference on AC2
Open-weight models on AC2 inference
Route to models like Qwen, GLM, Kimi, Nemotron, and Beam on autoscaling AC2 deployments.
Models you trained in AC2
Promote a checkpoint from a training run to serving and add it to the router's eligible models.
One data loop
The same production traces that train the router can also train the models it routes to.
Online evals
Monitor router traffic with evaluators that run on a subset of traces and measure quality with a function or LLM call.
Train a router for your workload
AC2 Routers are in private beta. Reach out to onboard and learn more.