Scaledown team released CDMs, a family of constrained-decoding models including Classify, Extract, and Summarize, each specialized for a single task while skipping token-by-token generation. They support 1M multimodal input tokens, score 77.2% on DecisionBench, are claimed to be 200x faster and 467x cheaper than frontier LLM APIs, with input pricing at $0.05/M tokens and free output, and are already live on platforms like OpenRouter, Vercel, LiteLLM, and LangChain.