Introducing Shift-1, a frontier coding router trained on AC2 that reduced coding agent cost by up to 55% with no noticeable drop in quality. Try it at

Read more on how to train and deploy your own custom router.

Congratulations!

World class! Congrats @ypatil125 & team

Huge !

Router efficiency! This cuts the noise so we can build faster 🙂

Congrats!

💸💸💸

I love saving money this is great

feels like most of the router community talks about optimizing cheaper models but gloss over the KV cache switch cost :/. so glad y'all explicitly included it in your analysis :). how do the savings change between short and long sessions?

如果代码质量真能保持住，省下来的调用成本，对每天跑很多任务的团队就很实在。

yay

@jaredrosnerd Congrats @ypatil125

Great building with you on routing research. Congrats on Shift-1! 👏

Very cool. Pumped to use this.

send "rename this variable" to a model that doesn't need to contemplate existence first

"No noticeable drop in quality" is doing the router's PR for it. Noticeable to who, on what tasks, measured how? A router that's 55% cheaper and silently worse on the 10% hardest diffs is still a router that's silently worse.

Smarter model selection. Lower costs. No compromise on intelligence. Shift-1 brings request-level routing to agentic AI, cutting costs by up to 55% without sacrificing quality. Huge congratulations to Yash, Rhythm, Linden and the entire AC team on this breakthrough! 🚀

Great tech! but i feel like the real benefit is when you move from Sol to Luna. Also open source benchmarks (terminal bench or the likes) to prove ur claims would be great

@grok what is AC2 & compare it with other model routers who are it’s direct competitor

55% cheaper with no quality drop sounds good. I want to know where the savings come from. If Shift-1 handles easy turns cheaply and sends hard ones to the frontier model, thats smart routing. If it just gives shorter answers, thats a different tradeoff. How often does it keep a turn on the cheap model, and how does it recover when it gets the call wrong?