Found first: a primary source the press has not covered yet.
A COLM 2026 paper finds the best LLM and the best embedding model score within 0.4 points of each other across 37 text tasks. The LLM costs up to 1,431 times more to run. Adnan El Assadi, Niklas Muennighoff, and Jinhyuk Lee publish their full comparison on arXiv, covering 36 models and five task categories.
What the source says
The study pits ten LLMs across six families against 26 embedding models ranging from 118M to 14B parameters, across 37 tasks spanning classification, semantic textual similarity, clustering, pair classification, and retrieval. The top LLM, Gemini 3.1 Pro, scores 77.6; the top embedding model scores 77.2. The cost gap reaches up to 1,431 times, or USD 154 versus USD 0.11 per benchmark pass. Open LLMs run 2.5 to 736 times more slowly on the same hardware. Reasoning tokens account for 28 to 81% of LLM inference cost, and the paper finds that lower reasoning budgets preserve or improve retrieval quality.
Why it matters
This gives production teams the first controlled, cost-aware answer across this many tasks and models. LLMs hold an edge on reasoning-heavy retrieval; embedding models lead in classification. The Pareto frontier across all 37 tasks includes only leading embedding models and Gemini 3.1 Pro. Teams using LLMs for classification or clustering workloads are paying a cost premium the task does not require.