Text Embedding Models: Specs, Prices and Availability
Match the exact model ID and service before comparing embedding specifications. Cohere embed-v4.0 accepts 128k tokens and defaults to 1,536 dimensions. Google text-embedding-004 has different retirement dates in the Gemini API and Google Cloud. A model name alone does not identify an available endpoint or a current price.
Updated October 28, 2026. This reference checks eleven model versions already covered by this page, including older versions that readers may still have in an integration. It is not a list of the newest releases. For a current evaluation shortlist and workload-cost scenarios, use our embedding model selection guide.
Our data: eleven model records and calculated storage
We checked publisher documentation for eleven model IDs, separated their input limits, vector dimensions and deployment conditions, and calculated eleven raw-storage values for the same 100,000-vector workload. The storage formula and configuration assumptions appear below. This is an AI-assisted documentation review by PE Collective; no model inference, latency test or retrieval benchmark was performed.
The comparison preserves older model IDs so you can check an existing application without confusing a replacement model with a drop-in configuration change. Specifications were checked on the update date. Prices, endpoint availability and commercial terms need another check before a deployment decision.
What dimensions and input limits do these versions support?
Dimensions below are the documented default or full output size. An input limit describes the accepted sequence, not a recommended chunk length. Provider tokenizers differ, so the same document can consume different token counts.
| Exact model ID | Dimensions | Input limit | Primary documentation |
|---|---|---|---|
| text-embedding-3-small | 1,536 default | 8,192 tokens | OpenAI |
| text-embedding-3-large | 3,072 default | 8,192 tokens | OpenAI |
| embed-v4.0 | 1,536 default | 128k tokens | Cohere |
| voyage-3-large | 1,024 default | 32,000 tokens | Voyage |
| voyage-3-lite | 512 | 32,000 tokens | Voyage |
| jina-embeddings-v3 | 1,024 full size | 8,192 tokens | Jina model card |
| text-embedding-004 | 768 maximum | 2,048 tokens | Google Cloud |
| BAAI/bge-large-en-v1.5 | 1,024 | 512 tokens | BAAI model card |
| intfloat/e5-large-v2 | 1,024 | 512 tokens | E5 model card |
| Alibaba-NLP/gte-large-en-v1.5 | 1,024 | 8,192 tokens | GTE model card |
| nomic-embed-text-v1.5 | 768 full size | 8,192 tokens | Nomic model card |
OpenAI supports shorter output vectors through its dimensions parameter. Cohere v4 offers 256, 512, 1,024 and 1,536 dimensions. Voyage 3 large offers 256, 512, 1,024 and 2,048. Jina and Nomic document their own resizing procedures. Use the selected model's documented settings; equal output dimensions do not establish that two models share a vector space.
Which availability changes affect existing integrations?
Google: the Gemini API deprecation table lists January 14, 2026 as the shutdown date for text-embedding-004 and recommends gemini-embedding-2. Google's separate Cloud model lifecycle table lists April 1, 2027 as the retirement date for text-embedding-004. These describe different services. Check your API hostname, project, model ID and region before diagnosing a failed request or choosing a migration date.
Cohere: embed-v4.0 remains in the documentation, while Embed 5 Pro and Fast are presented as the latest family. The v4 limits in the table belong to v4.0; do not copy a newer family's default dimensions into an existing index. This page previously described v4 as having a 512-token limit. The current primary documentation supports 128k.
Voyage: the specifications list the 3-series under previous models and describe the 4-series separately. Keep an explicit version in application configuration. A newer model can change dimensions, pricing or retrieval behavior even when its name looks similar. Review the provider's compatibility guidance before reusing document vectors.
A downloaded model card establishes documented capabilities, not successful service access. Account permissions, regional support, quotas and serving configuration can still affect an integration. Record an actual endpoint check during your deployment process; this reference did not call the embedding APIs.
What prices can be compared on the same basis?
Four rates for these exact versions are published in the primary sources below. They are standard input-token rates, before credits, discounts or taxes. A hosted API price and the cost of operating downloaded weights are different budgets.
| Model | USD per million input tokens | Source |
|---|---|---|
| text-embedding-3-small | $0.02 | OpenAI pricing |
| text-embedding-3-large | $0.13 | OpenAI pricing |
| voyage-3-large | $0.18 | Voyage previous-model pricing |
| voyage-3-lite | $0.02 | Voyage previous-model pricing |
For Cohere, Jina and Google's Cloud endpoint, confirm a price for the exact service and deployment you will use. This review does not establish a comparable current token rate for those entries. It therefore withdraws the old unverified Jina value ranking and the combined Google “Free / $0.025” label. Missing verification does not mean a provider has no public price.
Calculate API expense as input tokens / 1,000,000 × rate. Include document refreshes, query embeddings and retries. Storage, indexing, reranking, answer generation and operations require separate estimates. There is no universal monthly bill or model-volume threshold at which self-hosting becomes cheaper; measure the workload and serving costs you actually expect.
How much raw vector storage does each configuration require?
For 100,000 dense vectors stored as float32, calculate decimal megabytes as 100,000 × dimensions × 4 bytes / 1,000,000. The table uses the default or full-size output listed above, with one vector per input. These eleven values are calculated payload sizes, not database capacity measurements.
| Model | Raw storage for 100,000 vectors |
|---|---|
| text-embedding-3-small | 614.4 MB |
| text-embedding-3-large | 1228.8 MB |
| embed-v4.0 | 614.4 MB |
| voyage-3-large | 409.6 MB |
| voyage-3-lite | 204.8 MB |
| jina-embeddings-v3 | 409.6 MB |
| text-embedding-004 | 307.2 MB |
| BAAI/bge-large-en-v1.5 | 409.6 MB |
| intfloat/e5-large-v2 | 409.6 MB |
| Alibaba-NLP/gte-large-en-v1.5 | 409.6 MB |
| nomic-embed-text-v1.5 | 307.2 MB |
The same dimension count produces the same raw payload even when the models behave differently. Index structures, text, metadata, replicas and backups add space. Quantization or shorter documented outputs change the estimate. Measure retrieval again after changing the representation rather than assuming the smaller index will preserve results.
What does the downloaded model license say?
The linked publisher cards label BGE-large-en-v1.5 and E5-large-v2 as MIT, and GTE-large-en-v1.5 and Nomic-embed-text-v1.5 as Apache 2.0. Jina's v3 card labels its weights CC BY-NC 4.0 and directs commercial-use inquiries to the provider. Check the actual license and deployment terms for the version you intend to use; downloadable weights do not by themselves establish unrestricted commercial use.
Budget for inference hardware, utilization, model loading, batching, memory and maintenance. A quoted GPU hourly price cannot establish embeddings per second without a specified model, input length and serving configuration. The old page's hardware-throughput and self-hosting break-even claims lacked that evidence and have been removed.
Which input settings need to stay paired?
Follow the model card's retrieval configuration for both indexed documents and queries. E5 documents query: and passage: prefixes. Nomic documents search_query: and search_document:. Jina exposes retrieval.query and retrieval.passage tasks. These settings are part of the model interface, not interchangeable naming conventions.
Save the exact model revision, dimensions, preprocessing, task instruction and normalization beside your index configuration. Inspect input lengths before calling the model so you can detect truncation. When a deployment changes, compare its outputs with a saved corpus and representative labeled queries before routing production searches to it.
Why is there no combined MTEB ranking?
The MTEB project evaluates embeddings across multiple tasks, languages and modalities. A score needs its benchmark version, task selection, metric, languages and model configuration. An average across different tasks cannot be substituted for an English retrieval result.
The previous table did not retain enough evidence to show that its scores were comparable. We removed the ranking and the claim that a headline average predicts production search performance. Use matched benchmark results to shortlist candidates, then evaluate relevant-passage coverage and ordering on your own data. The sibling selection guide describes that evaluation process and reproducible cost scenarios.
What should you verify before changing an existing index?
- Record the exact current model ID, endpoint and dimension count. Check the relevant provider's lifecycle page.
- Confirm that the replacement's license, service access and billing units fit the intended deployment.
- Re-encode a saved test corpus with documented query/document settings and inspect retrieval failures.
- Estimate the full re-embedding workload and temporary duplicate storage, then retain a rollback path.
PE Collective owns this documentation review and will recheck it when model lifecycle or pricing information changes. The immediate task for an existing integration is to identify its actual endpoint and version. Use those details to select the applicable row and migration guidance.