Workload
Distinct items in your corpus
Roughly 4 chars per token for English
How often you re-process everything
Each is a small embedding call
Results
| Provider · Model | Dimensions | Per 1M tokens | Backfill cost | Annual cost |
|---|
How pricing is calculated. Provider rates as of model launch documentation. Self-hosted models (BGE, Nomic local) show $0 token cost but you pay for GPU compute; estimate separately. Total = backfill (corpus tokens × frequency) + queries (query tokens × 12 months).
Notes on choosing
| Multilingual corpus? | Voyage multilingual or Cohere multilingual-v3 outperform OpenAI on non-English. |
| Long documents? | Voyage-3-large handles 32k context. OpenAI text-embedding-3 caps at 8k. |
| Self-hosted budget? | BGE-M3 and Nomic Embed v1.5 are competitive open-weight options. Pay GPU, not tokens. |
| Latency sensitive? | Smaller dims = faster search. text-embedding-3-small (1536d) or Jina v3 (1024d). |