Understudy Labs

About Understudy Labs

Understudy watches a frontier model do your production work, then trains a smaller open-weight model to do the same job for a fraction of the cost. The loop has four steps. Capture takes traces from real LLM workflows via a single install that deploys inside the coding agents you already use, with hosted infrastructure optional. Evaluate grades those traces and freezes a benchmark, so any future model swap has to meet or beat it in A/B testing rather than being adopted on vibes. Train fine-tunes a new model on prompts and weights you own outright, starting locally and scaling to cloud training as results justify it. Deploy ships the successor only after it beats a held-out eval, and production data feeds back into training so performance compounds. What Understudy optimizes is broader than model choice — the company describes it as the whole production route: harness, prompts, schemas, tool-call adapters, reasoning mode, token caps, scorers, retry policy, batching, context compaction, parsers, fine-tuned descendants, and serving path. The published numbers are the sales pitch: an Understudy ladder scoring 0.630 against Sonnet 4.6's 0.557 at roughly a quarter of the cost, a tuned Qwen3-8B matching Sonnet quality at 5.2x lower latency and 6x lower token cost, and a post-trained 30B model labelling 39,962 comments at 50x lower cost than Opus. You keep the weights, and nothing has to move into a hosted app — the CLI and MCP server work where your workflow already lives.

Key Features

Understudy Labs Pros & Cons

Who Is Understudy Labs Best For?

Tags