PostHog reports decision models improve LLM costs

PostHog shares practical lessons from experimenting with decision models in production, finding them up to 444 times cheaper and 193 times faster than frontier LLMs. The author explains how the team uses decision models to gate expensive LLM calls, act as model routers, perform face-value checks, and replace overqualified language models. However, the author notes that they are unsuited for multi-step reasoning or open-ended tasks and require per-task threshold tuning. Open article →