OpenAI text-embedding-3 iconOpenAI text-embedding-3

OpenAI's embedding API: text-embedding-3-small at $0.02 and 3-large at $0.13 per 1M tokens, 8,192-token input, shortenable vectors

$0.02 Per 1M tokens (3-small)
3,072 Max dimensions (3-large)
8,192 Max input tokens

Overview

OpenAI's embedding API turns text into vectors for search, clustering, recommendations and RAG. Two current models cover most needs: text-embedding-3-small (1,536 dimensions, $0.02 per 1M tokens) and text-embedding-3-large (3,072 dimensions, $0.13 per 1M tokens, which OpenAI calls its most capable model for English and non-English tasks). Both accept up to 8,192 input tokens and take a dimensions parameter that shortens the vector, so you can trade a little accuracy for much smaller storage; OpenAI says a 256-dimension 3-large vector still beats the older text-embedding-ada-002 at 1,536. As of November 2026 the pricing page lists no newer OpenAI embedding model, and ada-002 ($0.10) is still callable and not marked deprecated. Also from OpenAI: OpenAI, the company's model platform and API.

The Verdict

Who Should Use OpenAI text-embedding-3?

Best For

  • Teams already on the OpenAI API who want one vendor and one key
  • Cheap, solid English retrieval at scale with 3-small
  • Projects that need the widest SDK, framework and vector DB support

Not Ideal For

  • Long documents: input stops at 8,192 tokens (Voyage and Cohere take 32K to 128K)
  • Self-hosting or air-gapped use: weights are not available
  • Teams chasing top retrieval scores; newer rivals lead recent tests

What's Great

  • 3-small is among the cheapest hosted embeddings at $0.02 per 1M tokens
  • Shortenable vectors cut storage without retraining
  • Batch endpoint supported for large backfills
  • Stable model line: no forced migration since the 3 series launched

Watch Out For

  • No new embedding model in the current lineup; rivals shipped several generations
  • In a July 2026 hands-on RAG test, 3-large scored Recall@10 0.86 vs 0.89 for voyage-4-large; cutting it to 1,024 dimensions dropped it to 0.83
  • One 2026 blog ranking placed 3-large 13th of 15 on a retrieval board
  • Text only: no image or PDF embeddings

Pricing

View all features & details

Key Features

  • dimensions parameter on both v3 models (not on ada-002)
  • 8,192-token input limit on all three models
  • Embeddings and Batch endpoints
  • OpenAI-reported MTEB average: 3-large 64.6, 3-small 62.3, ada-002 61.0

Rate Limits (3-large)

  • Free tier: 100 requests and 40,000 tokens per minute
  • Paid tiers scale up to 10,000 requests and 10M tokens per minute
  • Batch queue limits rise with each usage tier

Use Cases

  • Semantic search and RAG retrieval
  • Clustering, classification and deduplication
  • Recommendations and anomaly detection

Real-World Usage

Independent Test (2026)

  • About 10,000 English technical docs, 120 labeled queries
  • Recall@10: voyage-4-large 0.89, 3-large 0.86, Cohere embed-v4.0 0.85
  • Adding a reranker gained about 6 points, more than the gap between embedders
  • Author's summary: OpenAI wins on ecosystem and stability

Lifecycle

  • First-generation OpenAI embeddings were shut down on 2024-01-04
  • ada-002 still listed and priced, with no deprecation date
  • Docs now live at developers.openai.com (old platform.openai.com links redirect)

How It Compares

Feature OpenAI 3-large Voyage AI Cohere Embed BAAI BGE
Newest model text-embedding-3-large voyage-4-large embed-v5.0-pro bge-m3
Max input 8,192 tokens 32K tokens 128K tokens 8,192 tokens
Price per 1M $0.13 (3-small $0.02) $0.12 (4-lite $0.02) Not published Free to self-host
Images No Multimodal model Yes Separate BGE-VL
Open weights No voyage-4-nano only No Yes (MIT)

Sources: OpenAI pricing · Voyage pricing · Cohere models · bge-m3 card

Is this your tool? Add a badge