A cost-aware evaluation compared large language models against specialized embedding models across 37 text-understanding tasks and found the best LLM and the best embedding model achieved nearly identical performance scores. Despite that parity, the analysis found LLMs can cost up to 1,431 times more than a comparable-quality embedding model to run. The study recommends using embedding models for similarity and classification tasks while reserving LLMs for reasoning-intensive retrieval work where their extra cost is actually justified.
