Data · Updated Aug 16, 2026

Embedding Model Price Index 2026

Every major embedding API's list price per 1M tokens in one table — from OpenAI text-embedding-3-small at $0.02 to free open-source bge models you self-host. Tracked weekly, free forever.

Last verified 2026-08-16 · 12 models · 9 providers
$0.02
Cheapest API /1M
$0.00
Open-source tier
3072
Max dimensions
12
Models tracked

Complete embedding price table

List prices in USD per 1M tokens. Open-source models cost $0 in API fees when self-hosted. Prices verified 2026-08-16.

Model Provider Dimensions Context Price $/1M Best for
text-embedding-3-largeOpenAI30728191$0.13Maximum retrieval quality; RAG at scale
voyage-3Voyage AI102432000$0.12Long documents, code retrieval
text-embedding-ada-002OpenAI15368191$0.10Legacy; still the most deployed model
cohere-embed-v3Cohere1024512$0.10Enterprise search and RAG
mistral-embedMistral10248000$0.10EU hosting, RAG pipelines
jina-embeddings-v3Jina AI10248192$0.0689-language multilingual
gemini-embeddingGoogle30722048$0.05Multimodal + text retrieval
qwen3-embeddingAlibaba102432768$0.04Long-context, cost-sensitive
text-embedding-3-small cheapest APIOpenAI15368191$0.02Best price-to-quality ratio
bge-largeBAAI · open source1024512$0.00self-host Local RAG, privacy-first
bge-m3BAAI · open source10248192$0.00self-host Multi-lingual, multi-granularity
gte-largeAlibaba · open source10248192$0.00self-host Strong zh/en retrieval

Dimensions vs cost: the 1536 / 768 / 384 tradeoff

Output dimension is the hidden line item of any embedding budget. It does not change the API price, but it multiplies your vector database bill, query latency and memory footprint.

Dimensions Storage / 1M vectors (float32) Storage / 1M vectors (int8) Typical use
3841.5 GB0.4 GBMobile, latency-critical, budget search
7683.0 GB0.8 GBDefault RAG balance
15366.0 GB1.5 GBtext-embedding-3-small / ada-002 default
307212.0 GB3.0 GBtext-embedding-3-large / Gemini max

How to read this

  • 4× spread: a 3072-dim corpus costs 4× more to store than a 768-dim corpus, and HNSW/IVF index builds run roughly 1.5–2× slower at high recall targets.
  • Matryoshka truncation: text-embedding-3 models can be truncated to 512/256 dims at embedding time — pay for quality once, serve a smaller index.
  • Quantization: int8 (shown above) cuts storage 4× with recall loss typically under 1 point at 768+ dims; binary quantization (1 bit) works surprisingly well above 1024 dims.
  • Practical rule: start at 768 dims for most RAG workloads. Only step up to 1536+ when retrieval quality benchmarks (nDCG@10 on your own data) show a real gain.
  • Context matters too: 512-token contexts (bge-large, Cohere) force chunking of long documents; 8192+ contexts (bge-m3, qwen3-embedding, voyage-3) let you embed whole sections at once — fewer vectors, cheaper search.

Cost calculation example: 1M documents × 500 tokens

A typical knowledge-base migration: one million documents averaging 500 tokens each = 500M tokens total. Here is what each model charges for the one-time embedding pass.

Model Rate $/1M 500M tokens One-time cost
text-embedding-3-small$0.02500 × $0.02$10
qwen3-embedding$0.04500 × $0.04$20
jina-embeddings-v3$0.06500 × $0.06$30
text-embedding-ada-002$0.10500 × $0.10$50
cohere-embed-v3$0.10500 × $0.10$50
voyage-3$0.12500 × $0.12$60
text-embedding-3-large$0.13500 × $0.13$65
bge-large (self-hosted)$0.00—$0 + GPU

Don't forget the storage line

  • Vector index: 1M vectors at 1536 dims ≈ 6 GB float32 — roughly $6–12/month on managed vector databases, vs $2–3/month at 384 dims.
  • Re-embedding: every model change re-runs the full pass. At 500M tokens, switching from text-embedding-3-large to text-embedding-3-small saves $55 per migration.
  • Incremental loads: 100K new docs/month at $0.02/1M ≈ $1/month — embedding cost is almost never the bottleneck once the index exists.
  • DrAI note: all of these models are available through DrAI's unified API at one flat subscription — see pricing.

Open-source vs hosted embedding APIs

The $0.00 rows above are not free — they trade API fees for engineering time and hardware. Here is the honest comparison.

When self-hosting wins

  • Zero per-token metering: bge-large on one GPU comfortably embeds ~10M documents; the API bill for the same volume would run $1,000+ at hosted rates.
  • Data stays in your VPC: regulated industries (healthcare, finance, EU public sector) often cannot send documents to third-party embedding APIs.
  • Break-even math: a $150/month GPU node replaces hosted embedding at roughly 2–5M documents/month of sustained traffic — below that, hosted wins on total cost.

When hosted wins

  • Quality ceiling: frontier hosted models (text-embedding-3-large, voyage-3) still beat open weights on long-tail and multilingual retrieval benchmarks.
  • Zero ops: no model updates, GPU monitoring, or index-rebuild duties; hosted models are re-trained and swapped in by the provider.
  • Batch throughput: 500M-token one-time migrations finish in hours on hosted APIs instead of days on a single GPU.
  • Hybrid play: most production teams embed with a hosted model and fall back to open weights for bulk offline jobs — the cost curve bends either way.

Embed anything. One flat price.

Stop metering tokens. DrAI bundles frontier LLMs, embedding models and image generation into a single subscription with one unified API — no per-token bills, no surprise invoices.

See DrAI pricing Start free — no card needed
🌐 English ▾