ScaleDown ships Constrained Decoding Models: 1M-token task models for summarize, compress, extract and classify, with LangChain Deep Agents middleware

ScaleDown's Constrained Decoding Models (CDMs) target the high-volume context work around agent loops: summarizing history, compressing tool results, extracting exact fields and classifying state. Unlike a general LLM that writes any string token by token, each CDM is trained for one task and can only return the answers that task allows, such as a label from a caller-defined list or spans of the input, which lets it be much smaller and faster. Four models exist (Summarize, Compress, Extract, Classify), each accepting up to 1M input tokens; Extract, Summarize and Classify also take images and scanned PDFs. ScaleDown reports 200x faster and 467x cheaper than frontier models on these tasks, and versus the average of GPT-5.5, GPT-5.4 Mini and GPT-5.4 Nano, Summarize is 41x cheaper and 2.7x faster while Compress is 138x cheaper and 1.8x faster, both with higher accuracy (vendor-reported, no independent replication yet). LangChain released langchain-scaledown (0.2.0, MIT) with three middlewares for Deep Agents and create_agent, plus LangSmith LLM Gateway support at the /scaledown path. Extract and Summarize are on OpenRouter at $0.04 per million input tokens; Classify-1 runs on OpenRouter's Decisions API and is request-compatible with Jev. Keep a general-purpose LLM for open-ended generation like writing code and planning.