OpenAI text-embedding-3-large API Pricing & Vector Dimension Guide (2026)

Matryoshka Dimension Truncation Math & Production Vector DB Sizing

OpenAI's `text-embedding-3-large` operates natively at 3,072 dimensions, producing high-resolution semantic representations across 8,191 token context windows. However, indexing raw 3,072-dimensional float32 vectors into production vector databases introduces significant RAM and disk overhead: each vector consumes 12.288 KB of memory. By leveraging Matryoshka Representation Learning (MRL), developers can truncate vectors down to 1,536 or 1,024 dimensions via the `dimensions` API parameter, slashing vector database infrastructure costs by up to 66% while retaining over 98.5% of MTEB retrieval performance.

Production Asynchronous Embedding Ingestion in Python

Batch document chunks and enforce truncated dimensions directly at the API call layer:

from openai import AsyncOpenAI
import numpy as np

client = AsyncOpenAI()

async def embed_documents_mrl(texts: list[str], target_dims: int = 1536):
    response = await client.embeddings.create(
        input=texts,
        model="text-embedding-3-large",
        dimensions=target_dims
    )
    embeddings = [item.embedding for item in response.data]
    return np.array(embeddings, dtype=np.float32)

Cosine Similarity Invariance & L2 Normalization

When utilizing the OpenAI `dimensions` API parameter, the API server automatically truncates the high-dimensional projection and performs an L2-normalization pass on the resulting vector. Consequently, dot product search operations ($A \cdot B$) on these vectors are mathematically equivalent to cosine similarity, enabling database query engines to utilize fast SIMD inner product kernels without redundant normalization passes.