AI Potluck
Back to Gap Map Model components / Embeddings & retrieval models

OpenAI Text Embeddings

OpenAI
closed / Overall score: 3.3

OpenAI's general-purpose text embedding models, text-embedding-3-small and text-embedding-3-large, reached only through the /v1/embeddings endpoint. Both accept 8,192 tokens, default to 1,536 and 3,072 dimensions, and were trained with Matryoshka-style shortening so a caller can trade dimensions for storage through the `dimensions` parameter.

Still the January 2024 generation as of 2026-09-11: OpenAI's own embeddings guide lists text-embedding-3-small, text-embedding-3-large and the legacy text-embedding-ada-002 and nothing newer, so there is no text-embedding-4. No weights, corpus or training code are published, which is why openness stops at the ladder's first rule. Verified 2026-09-11 via the OpenAI embeddings guide and the Azure AI Foundry model list.

Openness

1 high confidence
1.0
weights
closed(no checkpoint distributed
data
closed(training corpus neither released nor described)
code
closed(no training or inference implementation
license
proprietary(hosted API under OpenAI business terms

API-only. The guide documents an endpoint, a dimensions parameter and a price, and nothing that can be downloaded, so the ladder's first rule fires ahead of any license question.

  • https://developers.openai.com/api/docs/guides/embeddings recorded 2026-09-11

    Documents text-embedding-3-small and text-embedding-3-large as endpoint calls with an 8,192-token limit and a `dimensions` parameter; the page offers no weight download, no corpus description and no training code, and names no successor model.

Adoption

4 medium confidence
4.0

No artifact to count and no call volume published. The level rests on distribution and on the models' standing as the industry's default baseline: Microsoft resells both SKUs directly through Azure AI Foundry, and every rival's evaluation benchmarks against OpenAI v3 Large as the reference point rather than against each other. Directional, so confidence stops at medium.

Capability

3 high confidence
3.0

Rung 3, solid workhorse: a superseded generation still carrying credible published numbers and enormous everyday use. It is two and a half years old, short-context by 2026 standards at 8,192 tokens, text-only, and the largest measured gap in Voyage's own RTEB comparison. That is the rung's definition rather than a demotion.

  • https://developers.openai.com/api/docs/guides/embeddings recorded 2026-09-11

    Model table: text-embedding-3-small 62.3%, text-embedding-3-large 64.6%, text-embedding-ada-002 61.0% on 'Performance on MTEB eval', max input 8192 for all three.

  • https://blog.voyageai.com/2026/01/15/voyage-4/ recorded 2026-09-11

    'voyage-4-large is the top-performing model, surpassing voyage-4, voyage-4-lite, Gemini Embedding 001, Cohere Embed v4, and OpenAI v3 Large by an average of 1.87%, 4.80%, 3.87%, 8.20%, and 14.05%, respectively' over all 29 RTEB datasets.

Verified 2026-09-11