The Embedder's Dilemma: LLMs Are Better, but at What Cost?

2026-08-13

AI Summary

This study systematically evaluates whether large language models (LLMs) can effectively replace dedicated embedding models in text embedding tasks, balancing performance against computational cost. For the first time, it quantifies the performance of 10 LLMs and 26 specialized embedding models across 37 tasks—including classification, semantic similarity, clustering, and retrieval—using a unified benchmark framework augmented with inference cost and latency analysis. Results show that both model families achieve comparable overall performance (77.6 vs. 77.2), yet LLMs incur, on average, 1,431× higher computational cost and are 2.5–736× slower. Only Gemini 1.5 Pro lies on the Pareto frontier. The findings demonstrate that optimal model selection is task-dependent, leading to a practical division-of-labor paradigm: embedding models are recommended for similarity and classification tasks, while LLMs should be reserved for retrieval scenarios demanding high reasoning capability.

Abstract

Should you replace your text-embedding pipeline with a large language model? We answer this with a controlled, cost-aware comparison of ten LLMs across six families and 26 embedding models (118M to 14B parameters) on 37 tasks spanning classification, semantic textual similarity (STS), clustering, pair classification, and retrieval. In aggregate the two paradigms are effectively tied: the best LLM (Gemini 3.1 Pro, 77.6) and the best embedding model (77.2) differ by 0.4 points. Their strengths differ by task: LLMs lead on reasoning-heavy retrieval, embedding models lead on classification, and the two match on clustering, STS, and pair classification. Reaching that parity is expensive. An LLM costs up to 1,431x more than an embedding model of comparable quality (USD 154 vs. USD 0.11 per benchmark pass), and the open LLMs tested process tokens 2.5 to 736x more slowly on the same GPU. Reasoning tokens account for 28 to 81% of LLM inference cost; lower reasoning budgets preserve or improve retrieval quality for most models in our ablation. The Pareto frontier contains the leading embedding models and one LLM, Gemini 3.1 Pro. These results support a division of labour: use embedding models for similarity, classification, and clustering, and reserve LLMs for reasoning-intensive retrieval. Our code, datasets, and results are publicly available at https://github.com/embeddings-benchmark/embedders-dilemma.

Problem

• large language models
• text embeddings
• cost-efficiency
• retrieval
• semantic similarity

Innovation

• text embedding
• large language models
• cost-efficiency
• retrieval
• Pareto frontier