Atlas Automated Embedding economics (Voyage tokens vs external embedding service)
Parent: Platform Consolidation TCO (Atlas Search/Vector Search vs Separate Search or Vector DB) · Published reference · snapshot 2026-10-30
↓ Facts as markdownall context files
Source-anchored research for Atlas Automated Embedding economics (Voyage tokens vs external embedding service). Run status: COMPLETED.
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
How it works
- Atlas regenerates embeddings when an indexed text field changes, but does not regenerate them for updates to fields outside Automated Embedding. Consequently, estimating re-embedding spend from every document update would overcount eligible updates. (low confidence) (verified as of 2026-10-14) [source]
- The operational consolidation benefit is that MongoDB manages embedding generation and updates and accepts query text for embedding generation, while charging per processed token. These sources describe that mechanism but do not measure engineering-hours savings or demonstrate a universal TCO reduction. (medium confidence) (verified as of 2026-10-14) [source]
Measurements and reference values
- Direct Voyage's pricing page advertises a 33% Batch API discount with a 12-hour completion window. Free token credits do not apply to batch usage; successfully processed batch tokens are billed and do not reduce the free balance. The Files API separately costs $0.05 per GB-month. (low confidence) (verified as of 2026-09-29) [source]
- Illustrative inference-only calculation: with a fresh 200-million-token voyage-4 allowance, 1 billion standard tokens cost (1000-200) x $0.06 = $48 on Atlas or direct Voyage. Direct Voyage batch would cost approximately 1000 x $0.06 x 0.67 = $40.20, before file storage. At 200 million tokens, standard inference is free while batch is approximately $8.04. This assumes identical token counts and the advertised 33% discount; it excludes queries beyond that workload, taxes, database/search infrastructure and engineering costs. (medium confidence) (verified as of 2026-10-14) [source]
Comparisons and alternatives
- Atlas documents a one-time 200-million-token allowance per model at organization level, shared across that organization's projects and clusters. It does not refresh, so a monthly TCO estimate must use the remaining balance rather than deducting 200 million every month. (medium confidence) (verified as of 2026-10-14) [source]
- Standalone Voyage's Batch API advertises a 12-hour completion window and a 33% discount versus standard endpoints. Free-token credits do not apply to Batch usage. A paid batch discount therefore does not imply batch is cheaper while standard requests remain covered by free tokens. (low confidence) (verified as of 2026-10-30) [source]
- OpenAI's model page lists text-embedding-3-small at $0.02 per million tokens and text-embedding-3-large at $0.13. Compared with Atlas's $0.02/$0.06/$0.12 Voyage 4 tiers, external service token pricing is model-dependent; these prices alone do not establish equivalent retrieval quality. (medium confidence) (verified as of 2026-10-30) [source]
- Cohere states that embedding models are priced by tokens embedded, and distinguishes billed units from generic token counts. For external-service cost reconciliation, use the provider's billed-unit fields rather than assuming every reported token count is chargeable. (low confidence) (verified as of 2026-10-29) [source]
- Doubleword reports an embedding experiment on 1,608,708 input tokens costing $0.03 for Qwen3 Embedding 8B batch. Its 100-query Wikipedia evaluation reports recall@10 of 82.4%, versus 85.1% for OpenAI's large model. This is a vendor-specific workflow demonstration, not evidence of equivalent quality or Atlas-wide TCO savings. (low confidence) (verified as of 2026-10-30) [source]
- Cohere documents embedding charges based on tokens embedded and describes trial API-key usage as free but limited. A comparison must distinguish trial benefits from production usage rather than treating a free trial as a recurring production allowance. (low confidence) (verified as of 2026-10-29) [source]
Facts and statements
- Atlas's published per-million-token prices are $0.02 for voyage-4-lite, $0.06 for voyage-4, and $0.12 for voyage-4-large; standalone Voyage lists the same standard endpoint prices. These lists show no Atlas token-price discount for those models. (medium confidence) (verified as of 2026-10-30) [source]
- Atlas meters embedding tokens during initial index creation, document inserts and updates, and queries. For indexing, only autoEmbed fields consume embedding tokens; query text also consumes tokens. Consolidating the pipeline therefore does not eliminate inference charges. (medium confidence) (verified as of 2026-10-14) [source]
- Cost-model inference: for a single Atlas voyage-4 model with its full 200-million-token grant remaining, 1 billion total processed tokens costs (1,000M - 200M) × $0.06/M = $48 in embedding fees. With the grant exhausted, the same volume costs $60. This excludes database, search, storage, tax, and engineering costs. (medium confidence) (verified as of 2026-10-14) [source]
- LangChain's MongoDBAtlasVectorSearch README configures OpenAIEmbeddings with text-embedding-3-large and stores/searches through Atlas. Thus, using an external embedding service does not by itself require choosing a separate vector database. (low confidence) (verified as of 2026-10-30) [source]
- MongoDB handles embedding generation, synchronization, and retries within its automated workflow. Consolidation can reduce application-owned pipeline work, but the reviewed sources do not quantify a transferable dollar saving for that engineering work. (medium confidence) (verified as of 2026-10-14) [source]
- Atlas stores generated embeddings in an internal collection on the same MongoDB cluster, with one generated collection per automated embedding index. Its cluster storage metrics exclude the Lucene vector index's disk use. Inference token prices alone therefore omit storage needed for the automated workflow. (low confidence) (verified as of 2026-10-14) [source]
- Atlas Automated Embedding and direct Voyage standard inference publish matching rates per million tokens for voyage-4-lite ($0.02), voyage-4 ($0.06), voyage-4-large ($0.12) and voyage-code-4 ($0.12). These list rates do not demonstrate an Atlas token-price discount over direct Voyage. (medium confidence) (verified as of 2026-10-14) [source]
- OpenAI lists text-embedding-3-small at $0.02 per million tokens, equal numerically to Atlas voyage-4-lite's list rate and one third of voyage-4's rate. This compares listed token prices for different models; it does not establish equal retrieval quality, equal tokenization or equal end-to-end cost. (medium confidence) (verified as of 2026-10-14) [source]
Children
- No children recorded.