Match each 2026 embedding vendor to a defining feature or pricing fact
Drag each answer to line up with its matching prompt
OpenAI text-embedding-3-small
Tightly paired with Cohere rerank-3.5 cross-encoder; popular for production RAG stacks
Voyage voyage-code-3
Open weights, multilingual (100+ langs), hybrid dense+sparse+multi-vector outputs
BGE-M3
Open weights, Matryoshka, designed for on prem and self-hosted deployments
Cohere embed-v3 / v4
Specialized for code retrieval, outperforms general models 10-30% on CodeSearchNet
Nomic Embed v1.5
$0.02/M tokens, 1536 dim (truncatable), default RAG choice in 2026
Each major 2026 embedding vendor owns a distinct positioning: OpenAI for default English, Voyage for code, BGE-M3 for multilingual hybrid, Cohere for paired rerank, Nomic for open-weights self-host.
Imagine a row of food trucks at a festival. Each truck specializes in one cuisine, and the menus do not really overlap. You go to the taco truck for tacos and the ramen truck for ramen. Search-tool providers work the same way. OpenAI is the general-purpose truck that serves the most customers. Voyage runs a code-specialist menu. BGE-M3 is the multilingual stand that speaks every language at the festival. Cohere comes paired with its own dessert, a matching ranker. Nomic is the take-home kit you cook in your own kitchen.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
The embedding-vendor market in 2026 has settled into five stable positions, each owned by a different specialist. Knowing the map collapses the model-selection problem from 'which of 40 options' to 'which of 5 buckets,' after which the choice within the bucket is largely a labeled-eval question.
This deep dive walks each of the five positions, the pricing or feature anchor that defines it, and why the positions are structural rather than likely to flip on the next vendor release.
Position 1: cost-optimized default
OpenAI text-embedding-3-small holds this position. At $0.02 per million tokens, it is roughly an order of magnitude cheaper than most credible alternatives. The Matryoshka truncation through the 'dimensions' parameter lets the same call serve as a 512-d, 1024-d, or 1536-d embedder, so storage tiers can choose the right dimension.
Why this position is stable
OpenAI has both the corpus access and the inference fleet to commodity-price embeddings. Newer vendors competing on price either lose money or compete on quality at a higher price point. The default position belongs to the vendor with the deepest fleet; that has been OpenAI for several years and shows no immediate sign of flipping.
When this is the answer
General English RAG, customer support, product description retrieval, blog and documentation search. Anywhere the corpus is plain English prose and the constraint is unit cost rather than peak quality.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
| Vendor / Model | Positioning | Hosting | Pricing or feature anchor |
|---|---|---|---|
| OpenAI text-embedding-3-small | default general English | API only | $0.02 / M tokens, 1536-d Matryoshka |
| Voyage voyage-code-3 | code retrieval specialist | API only | +10 to +30% on CSN over generalists |
| BGE-M3 | open-weights multilingual | self-hosted | 100+ langs, dense+sparse+multi-vector |
| Cohere embed-v3 / v4 | paired-reranker stack | API only | tight pairing with rerank-3 cross-encoder |
| Nomic Embed v1.5 | open-weights on-prem | self-hosted | Matryoshka, single-H100 deployable |
Real products, models, and research that use this idea.
- Cursor uses voyage-code-3 for repository-level code retrieval over user codebases.
- Many Databricks and Snowflake in-warehouse RAG stacks pick BGE-M3 or Arctic Embed to keep vectors inside the warehouse.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you validate the leaderboard claim that voyage-code-3 outperforms general embedders on code?
Build a small labeled retrieval set from your own codebase, embed and index with both, measure recall@10; the leaderboard is necessary but not sufficient.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Confusing voyage-3 with voyage-code-3, or claiming OpenAI is the cheapest without distinguishing 3-small from 3-large.
60 second bullets to scan on the way to the call.
OpenAI 3-small pricing and dimension
Voyage code specialist versus general Voyage SKU
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.