ScaleDown's Constrained Decoding Models are 200x faster and 467x cheaper than frontier LLM APIs
rohanpaul_ai · x · 2026-10-14
ScaleDown launched three small Constrained Decoding Models — Classify, Extract, Summarize — for the simple, repetitive tasks agents usually hand to expensive LLMs.
- Instead of writing answers token by token, CDMs do a single prefill pass over up to 1M tokens of text and images, then pick the answer straight from the input: a label, a span with source offset, or a phrase. Output is free and results stay inside your labels or source text.
- Priced at $0.05 per million input tokens, claimed 200x faster and 467x cheaper than frontier LLM APIs.
- The three models share one architecture: Classify returns calibrated probabilities with reasons (routing, tagging, intent), Extract emits well-formed JSON, Summarize pulls facts from long documents like due dates in 300-page PDFs.
More from coding & agent
- 歸藏 Declares War on Fake Accounts with an Auto-Block Plugin Batch-Flagging 70 Suspects — op7418 · 2026-10-14
- Writing a formal libtiff spec triggers CBRN checks: researcher slams safety gate false positive — jd_pressman · 2026-10-14
- McDonald's China opened an MCP, so this dev distilled your orders into a stomach skill — AlchainHust · 2026-10-14
- GovAI's Alan Chan on Governing AI Agents: Monitoring, Agent IDs and Rogue-Agent Risks — The Cognitive Revolution · 2026-10-14
- GovAI's Alan Chan on Governing AI Agents: Monitoring, Agent IDs and Rogue-Agent Risks — The Cognitive Revolution · 2026-10-14
- 6 key insights from Anthropic's skills guide: progressive disclosure is king — eherrerosj · 2026-10-14