What it is
text-embedding-3-large is OpenAI's most capable embedding model, released in January 2024 and still its newest as of October 2026. Embedding models turn text into vectors so software can find related passages for search, clustering, classification and retrieval-augmented generation (RAG).
It handles English and other languages, takes up to 8,192 tokens per input, and outputs 3,072 dimensions by default, which can be shortened with a dimensions parameter to save storage. It is text-only.
It costs $0.13 per million tokens through OpenAI's Embeddings API, with the Batch API at half price. The smaller text-embedding-3-small costs $0.02 per million.
Our review
text-embedding-3-large is the embedding model nobody is blamed for choosing, and the one most worth re-testing. OpenAI has shipped no new embedding model since January 2024, and its own launch figure is 64.6 on the original MTEB. Newer scores are often quoted against it unfairly: Qwen3-Embedding-8B's 75.22 is on MTEB English v2, a different benchmark, and Cognee's 2026 survey warns against ranking the two together. The cleaner comparison is Voyage's, which found voyage-4-large 14.05% better on average across the 29 datasets of the RTEB retrieval benchmark, though that is a vendor test.
Price is not where it loses. At $0.13 per million tokens, or half that through the Batch API, OpenAI estimates about 9,615 pages of 800 tokens per dollar, so embedding a ten-million-page archive costs roughly $1,040, or $520 in batch. Storage costs more than tokens: 3,072 float32 dimensions is 12 KB a vector, which is why the dimensions parameter matters, and OpenAI says the model cut to 256 dimensions still beats ada-002 at 1,536.
For enterprise buyers the draw is OpenAI's data handling. API inputs are not used for training unless a customer opts in, the embeddings endpoint keeps no application state, and it qualifies for Zero Data Retention without extra conditions; otherwise abuse-monitoring logs are kept for 30 days. Processing can stay in the US, Europe or, for this model alone, the UAE, though every region outside the US requires a signed retention amendment.
The limits are structural. Input stops at 8,192 tokens, so long documents must be chunked, and it is text-only while Cohere Embed 5 and Gemini Embedding 2 also take images. Moving off it later means re-embedding the whole corpus, so the real cost of choosing it is less the token price than the switching bill a team signs up for.
Where it earns its place
A RAG engineer at a 50-person SaaS company already using OpenAI
Embed the help centre and ticket history for a support bot
One API key and no new vendor review: 2M chunks of 500 tokens cost $130 to embed, $65 through Batch. Request 1,024 dimensions if the vector store caps there and accept a small loss in accuracy.
A compliance lead at a European insurer
Keep embedding traffic in EU processing with no retention
Use the eu. regional endpoint, which requires Zero Data Retention or modified abuse monitoring and a signed retention amendment. Because embeddings keep no application state, ZDR also removes the 30-day abuse log.
A search engineer weighing a migration for a 10M-document index
Decide whether to switch to a newer embedding model
Re-embedding about 8B tokens costs roughly $1,040, or $520 on Batch at up to 50,000 inputs per batch, so the token bill is small; re-indexing and evaluation are the real cost. Test voyage-4-large on your own queries, since the 14.05% gap is Voyage's measurement.
Before you commit
- Setup
- Call POST /v1/embeddings with model text-embedding-3-large and an optional dimensions value; any OpenAI SDK returns vectors within minutes of creating a key.
- Integrations
- Only the Embeddings and Batch endpoints are supported. Most vector databases and RAG frameworks have it as a built-in option.
- Data and privacy
- API data is not used for training unless you opt in; abuse-monitoring logs are kept 30 days unless Zero Data Retention is approved, which embeddings support. Regional processing in the US, Europe and the UAE; the data page names no SOC 2 or ISO certification.
- Cost at scale
- $0.13 per 1M tokens, $0.065 on Batch: a 10M-page archive at 800 tokens a page is about $1,040, or $520 in batch. Vector storage at 3,072 float32 dimensions is about 12 KB per chunk, usually the bigger bill.
- Limits
- 8,192 tokens per input; Batch allows 50,000 embedding inputs per batch with a 24-hour window. Rate limits run from 100 RPM and 40K TPM on Free to 10,000 RPM and 10M TPM on Grow; no deprecation date is announced.
What's good
- Supported by nearly every RAG framework and vector database
- $0.13 per 1M tokens, half that through Batch
- Zero Data Retention eligible; no training on API data by default
- dimensions parameter shrinks vectors from 3,072
What's not
- No new OpenAI embedding model since January 2024
- Voyage reports a 14.05% retrieval deficit against voyage-4-large
- 8,192-token input cap and text only
- Free tier limited to 40K tokens a minute
Best for
Teams already on OpenAI who value one vendor, strict data terms and broad tool support over top retrieval quality.
Alternatives
text-embedding-3-large: common questions
- Is text-embedding-3-large worth it?
- Safe, cheap and supported by every vector database, text-embedding-3-large now trails newer retrievers by enough to justify a re-test before any new build. We rate it 3.7 out of 5.
- How much does text-embedding-3-large cost?
- $0.13 per 1M tokens; Batch half price; text-embedding-3-small $0.02 per 1M.
- What are the downsides of text-embedding-3-large?
- No new OpenAI embedding model since January 2024; voyage reports a 14.05% retrieval deficit against voyage-4-large; 8,192-token input cap and text only; free tier limited to 40K tokens a minute.
- Who is text-embedding-3-large best for?
- Teams already on OpenAI who value one vendor, strict data terms and broad tool support over top retrieval quality.
- How much does text-embedding-3-large cost for a team or in production?
- $0.13 per 1M tokens, $0.065 on Batch: a 10M-page archive at 800 tokens a page is about $1,040, or $520 in batch. Vector storage at 3,072 float32 dimensions is about 12 KB per chunk, usually the bigger bill.
- Is text-embedding-3-large safe for company data?
- API data is not used for training unless you opt in; abuse-monitoring logs are kept 30 days unless Zero Data Retention is approved, which embeddings support. Regional processing in the US, Europe and the UAE; the data page names no SOC 2 or ISO certification.
- What are the best text-embedding-3-large alternatives?
- The closest alternatives we have reviewed are voyage-4-large, Gemini Embedding 2 and Cohere Embed 5.
Answers checked Oct 22, 2026.