Applied Compute released Shift-1, a coding router that picks a model for each request in an agent session instead of locking one model to an entire task. It works as a drop-in endpoint for Codex and Claude Code and was trained on AC2, the company's training and inference platform. On Applied Compute's internal coding usage, Shift-1 cut Codex spend by 55% compared with always using GPT-6-Astra, and cut Claude Code spend by 24% compared with always using Opus 5.5, with no discernible drop in intelligence. Coding agents often run for many turns, and a job that looks easy at the first message can get harder as the session continues. Later requests also cost more as context grows, so an early wrong choice is expensive. Shift-1's policy is a 4B parameter decision model that reassesses every request and accounts for the cost of switching models mid-session, including cache effects. The company is also releasing AC2 Routers so other teams can train and serve their own routers. Shift-1 is available to try at shift.ac. Applied Compute builds custom models for companies that want intelligence trained on their own data and workflows, then serves and improves those models from production traffic. The San Francisco company was founded in 2025 by former OpenAI researchers Yash Patil, Rhythm Garg, and Linden Li. AC2, the platform they used to train Shift-1, is in private beta.