OpenAI Embedding Models

A machine learning engineer at a legal tech startup needed to build a contract similarity engine. She tested three embedding models on 10,000 legal documents. text-embedding-3-small found similar contracts with 89% accuracy. text-embedding-3-large hit 94%. But when she used the dimensions parameter to reduce text-embedding-3-large from 3,072 dimensions to 1,024, accuracy only dropped to 92% while storage costs fell by 67%. That single parameter saved her company $14,000 per year in vector database hosting.

Choosing the right embedding model is a cost-accuracy-speed tradeoff. This lesson covers every option and helps you make the right choice.

The Three Models

text-embedding-3-small

The workhorse. 1,536 dimensions by default. The cheapest option and fast enough for real-time applications. This is the model you should start with for most projects.

MTEB benchmark score: 62.3% (average across retrieval, classification, and clustering tasks)

Best for: Most production applications, high-volume processing, applications where cost matters more than maximum accuracy, prototyping and experimentation.

text-embedding-3-large

The premium option. 3,072 dimensions by default -- twice the dimensionality of text-embedding-3-small, capturing more nuanced semantic relationships. Higher accuracy on difficult retrieval tasks where subtle differences in meaning matter.

MTEB benchmark score: 64.6%

Best for: Legal document comparison, medical literature search, academic research, any domain where the difference between "similar" and "nearly identical" matters. Also useful when you need the dimensions parameter to find a custom accuracy-cost tradeoff.

text-embedding-ada-002 (legacy)

The previous generation. 1,536 dimensions, no dimensions parameter for reduction. Still functional but outperformed by text-embedding-3-small at a lower price. No reason to use this for new projects.

MTEB benchmark score: 61.0%

Best for: Nothing new. Only relevant if you have an existing system built on ada-002 and migration is not worth the effort.

The Dimensions Parameter

This is the most powerful and underused feature of the text-embedding-3 models. You can request embeddings with fewer dimensions than the default, and the model produces a compressed vector that retains most of the semantic information.

from openai import OpenAI

client = OpenAI()

# Full 3072 dimensions
full = client.embeddings.create(
    model="text-embedding-3-large",
    input="Contract for sale of residential property"
)
print(len(full.data[0].embedding))  # 3072

# Reduced to 1024 dimensions
reduced = client.embeddings.create(
    model="text-embedding-3-large",
    input="Contract for sale of residential property",
    dimensions=1024
)
print(len(reduced.data[0].embedding))  # 1024

Why does this matter? Smaller vectors mean:

• Lower storage costs: A vector database storing 1 million 3,072-dimension vectors uses ~12 GB. At 1,024 dimensions, that drops to ~4 GB.
• Faster search: Comparing shorter vectors requires fewer mathematical operations.
• Lower memory usage: Smaller vectors fit in RAM, enabling faster indexing.

MTEB Scores: What They Mean

MTEB (Massive Text Embedding Benchmark) evaluates embedding models across dozens of tasks in multiple categories:

Token Limits and Pricing

Embedding models have a maximum input length of 8,192 tokens (roughly 6,000 words). Inputs longer than this are silently truncated -- the model processes only the first 8,192 tokens and ignores the rest.

import tiktoken

encoder = tiktoken.encoding_for_model("text-embedding-3-small")
tokens = encoder.encode(your_text)

if len(tokens) > 8192:
    print(f"Warning: text has {len(tokens)} tokens, will be truncated to 8192")
    # Consider chunking the text instead of letting it truncate

Pricing is per token, with batch discounts for large volumes. For cost estimation:

def estimate_embedding_cost(texts, model="text-embedding-3-small"):
    encoder = tiktoken.encoding_for_model(model)
    total_tokens = sum(len(encoder.encode(t)) for t in texts)
    
    # Approximate costs per 1M tokens (check current pricing)
    costs = {
        "text-embedding-3-small": 0.02,
        "text-embedding-3-large": 0.13,
        "text-embedding-ada-002": 0.10
    }
    
    cost = (total_tokens / 1_000_000) * costs[model]
    return {"total_tokens": total_tokens, "estimated_cost": f"${cost:.4f}"}