text-embedding-3-small API pricing

Estimate your embedding cost

Token count is measured from the text you send to the embeddings endpoint. These examples show only this model's embedding requests.

Word counts use the rule of thumb of about 0.75 English words per token; the tokenizer count is what is billed.

text-embedding-3-small vs 3-large vs ada-002

All three OpenAI embedding models are available through the same Crazyrouter endpoint. Switching models means re-embedding every document, so compare price, vector size and retrieval quality before you build the index. For providers beyond OpenAI, see the embedding model comparison.

Example: find a help article by meaning

A visitor asks “Where can I see how many tokens I used today?” Your knowledge base contains “View daily token usage in the usage dashboard.” The words differ, but embeddings place both sentences close together, so a vector search finds the right article.

Try the Python semantic-search example below to calculate the ranking yourself. Relevance depends on your content and queries; similarity scores are not confidence percentages.

From document count to project budget

Estimate the first indexing run, ongoing monthly embeddings and raw vector storage. Start with a sample workload or enter your own assumptions. No API key is needed for the calculator.

Copy an API example and start building

Use the OpenAI-compatible embeddings endpoint with the exact model ID. Create an API key and add funds for paid API requests. These examples run on your machine; this page never asks for your key.

curl --fail-with-body https://api.crazyrouter.com/v1/embeddings \
  -H "Authorization: Bearer $CRAZYROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "text-embedding-3-small",
    "input": "How do I reset my API key?",
    "encoding_format": "float"
  }'

Pricing questions

How much does text-embedding-3-small cost on Crazyrouter?

The current Crazyrouter rate is $0.02 per 1 million input tokens. The live pricing table is the source of truth.

How do I estimate the cost of a batch?

Multiply the input token count by the rate per 1 million tokens, then divide by 1 million. The estimate does not include any other models in your workflow.

Is there an output-token charge?

Embedding calls return vectors, not generated text. This page uses the input-token rate for the model; it does not apply a chat completion rate to the returned vector.

Should I use text-embedding-3-small or text-embedding-3-large?

Start with text-embedding-3-small for most search and RAG indexes: it costs less per token and stores smaller vectors. Try text-embedding-3-large when an evaluation on your own queries shows retrieval quality matters more than cost. Both models need a full re-index if you switch, because vectors from different models are not comparable.

Do I pay to embed my entire knowledge base for every search?

Store document vectors after indexing and reuse them. Each new search normally embeds only the query. Re-embed documents when their content changes, or when you change the embedding model or dimensions. The monthly calculator separates these costs.

Does reducing dimensions lower the API cost?

This estimate uses the same input-token rate at every dimension setting. Smaller vectors use less raw storage, but may change retrieval quality. Use matching dimensions for documents and queries, and verify the parameter with your API route before migrating an index.

Why does chunk overlap increase the estimate?

Overlap repeats part of a document in consecutive chunks. That repeated text is sent to the API again and counts toward input tokens. The planner includes the repeated tokens and does not pad the final chunk.

Does the estimate include RAG answer generation?

No. It covers embedding the documents and queries. A RAG workflow may also pay for a vector database, reranking, and a chat model to generate an answer from retrieved passages. Budget for those services separately.

How do I get an exact token count?

Use a tokenizer compatible with the model, such as tiktoken with cl100k_base, before a large run. Check usage.prompt_tokens in completed embedding responses. Character counts and average document lengths are planning assumptions, not exact billing counts.