An embedding model decides what your retrieval system can find. That is the framing OpenRouter uses in its 2026 embedding model guide, published 23 September 2026, and it is the right place to start. The model turns each input into a vector and places related inputs near one another, so retrieval works by meaning rather than exact wording. Change the model, and you change what your index can surface.
The shortlist is organized by input type
OpenRouter shortlisted models for English RAG, multilingual retrieval, code search, text-and-image retrieval, and low-cost indexing, then sent live requests through its embeddings endpoint. Its catalog returned 37 entries on 11 September 2026, including batch and preview variants. The guide’s own caveat matters: those API checks confirm request and response behavior, not retrieval quality.
For default English RAG, the recommendation is openai/text-embedding-3-small — low prompt price, an 8,192-token context, and adjustable output dimensions. When inputs exceed 8,192 tokens, or when you want to move between Voyage 4 tiers without rebuilding the index, the guide points to voyageai/voyage-4-large with a 32,000-token context and four selectable dimensions.
For multilingual retrieval with public weights, qwen/qwen3-embedding-8b is the pick: support for more than 100 languages, a 32,768-token context window, and downloadable weights. Code search goes to voyageai/voyage-code-4. Text-and-image retrieval goes to google/gemini-embedding-2, with voyageai/voyage-multimodal-3.5 as the second verified option.
Dimensions and price are index-size decisions
The dimensions column is where this stops being a leaderboard question. A 4,096-value vector uses four times the raw storage of a 1,024-value vector at the same numeric format. OpenRouter notes that openai/text-embedding-3-small defaults to 1,536 values, half of openai/text-embedding-3-large at 3,072, so Small uses half as much raw vector storage. You can shorten vectors with the dimensions parameter — a request with "dimensions": 256 returned 256 values — but smaller vectors can also reduce retrieval quality. Test before changing an existing index.
Price spreads are wide. perplexity/pplx-embed-v1-0.6b has the lowest prompt price among the paid text models in the shortlist at $0.004 per million input tokens, with a 32,000-token context. nvidia/nemotron-3-embed-1b:free is a free text route with a 32,768-token context window. The three free routes returned a 404 from OpenRouter’s test account because its privacy settings do not allow routing to providers that may train on free-model prompts; the default dimensions listed for those models come from their model cards, not from responses.
Compatibility changes the migration cost
Voyage states that all embeddings created with the 4 series are compatible with each other, which lets you test another Voyage 4 tier against an existing index. OpenRouter is careful here: a compatible vector space removes the rebuild requirement, but it does not guarantee identical retrieval results. Validate a tier change on a representative sample first.
Context limits can also be per endpoint rather than per model. The Qwen3 Embedding 8B model page lists 32,768 tokens, but on 11 September 2026 the DeepInfra and SiliconFlow endpoints listed 32,768 while Nebius listed 32,000. If you send longer inputs, pin a provider with the larger limit or keep inputs at or under 32,000 tokens.
What to do before you rebuild
OpenRouter’s method was to match each use case to a documented capability, then narrow by language coverage, context length, output dimensions, public weights, and prompt price. Published evaluations including MTEB and CoIR helped identify models to test. The Qwen model card, for instance, reports a multilingual MTEB score of 70.58 as of 5 June 2025 — a vendor-reported result to shortlist with, not a verdict.
The practical instruction is to compare at least two candidates on labeled queries and documents from your own application, and track how often each model retrieves the relevant chunks before building or rebuilding a large index. That is the same discipline behind OpenRouter’s tool-calling loop guidance: define the stopping condition before you commit. Here, the stopping condition is your retrieval target, not a benchmark ranking.
One limitation worth naming: the supplied guide covers only part of the catalog. It also lists baai/bge-m3, mistralai/mistral-embed-2312, and mistralai/codestral-embed-2505 as untested alternatives, plus a set of 512-token open models from BAAI, E5, GTE, and Sentence Transformers at $0.005 to $0.01 per million tokens. Providers can add or remove routes and prices can change, so check the current model page before starting a large indexing job.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.