Official reports indicate that trimming text-embedding-3-large from 3072 dimensions to 256 dimensions results in only approximately 2–3% loss in retrieval accuracy on the MTEB benchmark.