Embedding models, compared

text-embedding-3-small (OpenAI) and Amazon Titan Text Embeddings V2 (Amazon) share the lowest published rate we track, at $0.02 per 1M input tokens. Cheapest per token is not cheapest per job: an embedding you store is a recurring index cost, so the vector's dimension count drives your database bill long after the one-off embedding call is paid. Compare the price and the dimensions column together.

Because an embedding model does not return tokens — it returns a vector of floating-point numbers. Every provider here bills input tokens only, which is why this page has one price column instead of the input/output/blended trio used on the rest of the site. A blank in the price column means something different: the provider publishes no rate at all, and we leave it null rather than carry a figure we could not source.

Usually not. The number in the dimensions column is the model's default, and most of these families let the caller truncate it at request time — OpenAI via the `dimensions` parameter, Google via `output_dimensionality`, Mistral via `output_dimension`, and Cohere, Amazon Titan V2 and Amazon Nova 2 each via a fixed menu of sizes. Alibaba's Qwen v4 exposes eight steps from 64 to 2048. Each model page records the range its provider publishes, with the source it was read from. The input ceiling is the field you genuinely cannot change.

Far less than a chat model, and the spread is wide: Embed English v3.0 caps at 512 tokens while Embed 4 (embed-v4.0) takes 128K — and both are Cohere models, so the gap exists inside a single vendor's lineup. This is what decides your chunking strategy, and it is the field most price comparisons omit entirely.

4 of the 13 rows carry no rate because their provider does not publish one on any first-party surface. Cohere's pricing page lists only its current-generation embedding model, and the older family is absent from it and from the AWS Bedrock Price List API too. Aggregators quote a figure for those models; we do not repeat it. A null here means "the provider publishes nothing", which is a fact you can act on — a guessed price is not.