# text-embedding-3-small

text-embedding-3-small costs less than ada-002, scores higher on MTEB, and outputs the same default 1,536-dimension vectors. It's the straightforward upgrade for teams still running first-generation OpenAI embeddings and the sensible default for new projects that don't need the large variant's accuracy ceiling.

- **Model ID:** `openai/text-embedding-3-small`
- **Type:** embedding
- **Providers:** azure, openai
- **Context window:** Not applicable
- **Maximum output tokens:** 0
- **Pricing:** $0.02/1M input tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/text-embedding-3-small

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { embed } from 'ai';

const result = await embed({
  model: 'openai/text-embedding-3-small',
  value: 'Sunny day at the beach',
})
```

## About

OpenAI announced text-embedding-3-small on January 25, 2024 alongside its larger sibling. At reduced per-token cost compared to ada-002, it scores 62.3% on MTEB (Massive Text Embedding Benchmark), 1.3 points higher than the model it replaces.

Multilingual retrieval improves substantially too. text-embedding-3-small reaches 44.0% on MIRACL versus ada-002's 31.4%, a 12.6-point gain that matters for any pipeline handling non-English or mixed-language content.

Like text-embedding-3-large, text-embedding-3-small supports the `dimensions` parameter via Matryoshka training. The default 1,536-dimension output matches ada-002 for drop-in compatibility, but you can reduce it when memory or storage costs are a concern. Semantic structure is front-loaded into the earlier dimensions, so shorter vectors still carry meaningful signal.

text-embedding-3-small fits well in the query-time embedding path of Retrieval-Augmented Generation (RAG) architectures. Every user query must be embedded before retrieval, and at scale that per-query cost and latency compounds. Low cost and fast inference make it a natural fit for that leg of the pipeline.

## What to consider

Because the default output is 1536 dimensions, identical to ada-002, you can swap models without touching your vector database schema or index configuration.

## When to use

### Best For

- Drop-in replacement for ada-002 with no schema changes and immediate cost savings
- Query-time embedding in real-time RAG pipelines where latency and cost per request matter
- High-volume document indexing where the reduced cost yields significant infrastructure savings
- Multilingual retrieval with the 12.6-point MIRACL improvement over ada-002
- Budget-conscious projects where embedding quality is important but not the absolute ceiling

### Consider Alternatives When

- Retrieval accuracy is the bottleneck and the 2.3-point MTEB gap versus text-embedding-3-large is material for your use case
- Your corpus is heavily multilingual and the larger model's 54.9% MIRACL score (vs 44.0%) would produce meaningfully better results
- You need the maximum possible embedding dimensionality (3072) for specialized downstream models

## Best for

- **Ada-002 drop-in replacement:** No schema changes and immediate cost savings with the same 1536-dimension default
- **Real-time RAG pipelines:** Query-time embedding where latency and cost per request matter
- **High-volume indexing:** Document pipelines where the reduced cost yields significant infrastructure savings
- **Multilingual retrieval:** Applications that benefit from the 12.6-point MIRACL improvement over ada-002
- **Budget-conscious projects:** Embedding quality is important but not the absolute ceiling

## Consider alternatives

- **Retrieval accuracy bottleneck:** The 2.3-point MTEB gap versus text-embedding-3-large is material for your use case
- **Heavily multilingual corpus:** The larger model's 54.9% MIRACL score versus 44.0% would produce meaningfully better results
- **Maximum dimensionality:** Specialized downstream models need the full 3072-dimension vectors from text-embedding-3-large

## Frequently asked questions

### Is text-embedding-3-small a direct replacement for ada-002?

Yes. The default output is 1536 dimensions, same as ada-002, so existing vector indexes work without rebuilding. You get a higher MTEB score and immediate cost savings.

### How much does the multilingual retrieval improve over ada-002?

MIRACL scores go from 31.4% to 44.0%. For pipelines that handle queries or documents in multiple languages, this is a meaningful quality improvement that comes free with the model swap.

### When does it make sense to pay for text-embedding-3-large instead?

When your application's quality is bottlenecked by embedding accuracy, for example, legal search, scientific literature retrieval, or high-stakes recommendation systems where a 2-point MTEB difference translates to noticeably better results.

### Can I reduce the vector dimensions below 1536?

Yes. The `dimensions` parameter accepts any value below the default. Matryoshka training ensures the truncated vectors retain useful semantic structure, which is helpful for reducing storage costs in large indexes.

### What are typical latency characteristics?

This page shows live throughput and time-to-first-token metrics measured across real AI Gateway embedding traffic.

## Links

- [Model page](https://vercel.com/ai-gateway/models/text-embedding-3-small)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://platform.openai.com/docs/models/text-embedding-3-small)
- [Provider pricing](https://platform.openai.com/docs/pricing)