Cost metrics estimate what each embedding model would cost you in production: indexing your documents and answering queries. They’re computed from the tokens the model processes and its price per 1 million input tokens, in US dollars.
Running evaluations in Truvec doesn’t cost you anything extra: your plan includes the usage. These prices are only there to compare models.
Supported models
Open-source models are priced at the rate of the host that runs them for Truvec (Fireworks AI). If you host one yourself, your cost depends on your hardware instead.
Each model is used the way its provider recommends for search: questions and documents are embedded with the matching input type or instruction (for example Cohere’s search_query / search_document, or Nomic’s search_query: prefix). Models that read at most 512 tokens (Cohere v3, BGE, mxbai, GTE) only see the beginning of longer chunks, exactly as they would in production.
Truvec keeps these at the providers’ list prices. You can see the current prices in Settings, under Embedding model pricing.
If you negotiated a discount with a provider, scale the cost figures in your results and reports by the same factor: the comparison between models stays valid.
When prices change
Costs are computed when a run happens. When a provider changes its price, runs made afterwards use the new price, while past results and reports keep the price used at the time, so earlier reports stay accurate. 
