OpenAI embeddings vs a pay-per-call embeddings API with no account
Updated 2026-10-08
OpenAI embedding models are a strong, widely used default. Tanod's /v1/embed serves two small open models on our own CPU, paid per call with x402 and no account. The trade is simple: we need no sign-up, but we are smaller and, per token, much more expensive.
At a glance
| OpenAI embeddings | Tanod /v1/embed | |
|---|---|---|
| Model | text-embedding-3-small (default 1536 dimensions); text-embedding-3-large (3072). | BAAI/bge-small-en-v1.5 (English, MIT) or intfloat/multilingual-e5-small (about 100 languages, MIT). 384 dimensions. |
| Max input | 8192 tokens per input. | 512 tokens per text, cut and flagged; 64 texts per call; 16,384 tokens per request. |
| Quality (MTEB) | Its docs list 62.3% for 3-small and 64.6% for 3-large. | The model card lists an MTEB average of 62.17 over 56 datasets (model card). Different reporting, so not strictly comparable. |
| Vector size | 1536 floats by default; a dimensions parameter can shorten it. | 384 floats: about a quarter of the storage per vector. |
| Account | API account and key. | None. x402 payment in USDC, or the free tier. |
| Price | Per input token; about USD 0.02 per 1M tokens for 3-small (derived, see below). | USD 0.0005 per text, minimum USD 0.001 per call. |
| Data handling | See OpenAI's policies. | Text is processed in memory and not logged or stored. Runs on our own CPU; no third-party API. |
When to choose OpenAI embeddings
- You embed large volumes. At roughly USD 0.02 per 1M tokens it is far cheaper per token than ours.
- You need long inputs (up to 8192 tokens) or higher-dimensional vectors, or you need the larger model.
- You already use OpenAI, and your index is built on its vectors. Vectors from different models are not interchangeable.
When to choose Tanod
- An agent or script needs vectors now, without creating an account or holding an API key, or you want to try retrieval with a few calls.
- English text, or text in about 100 languages with the multilingual model, in small volumes, with small 384-dimension vectors.
- You prefer an open, MIT-licensed model that you can later run yourself, with the same vectors.
Worked example. 1,000 texts of about 200 tokens each is about 200,000 tokens. At the 3-small rate derived from OpenAI's "62,500 pages per dollar" at about 800 tokens a page, that is roughly USD 0.004. On ours it is 1,000 times USD 0.0005, so USD 0.50. For bulk jobs choose the provider with the lower token price, or run bge-small yourself.
The OpenAI price is our arithmetic from its published pages-per-dollar figure, not a quoted rate; check the current rate on its pricing page. We have not benchmarked the two on your data. Test retrieval on a sample of your own documents before choosing.
Pricing as of 2026-10-08
OpenAI (from its embeddings guide, 2026-10-08): priced per input token; 62,500 pages per dollar for 3-small and 9,615 for 3-large at about 800 tokens a page, i.e. roughly USD 0.02 and USD 0.13 per 1M tokens.
Tanod: USD 0.0005 per text, at least USD 0.001 per call (1 to 64 texts, so USD 0.001 to USD 0.032), paid in USDC on Base or Polygon. 5 free calls per IP per UTC day shared across the ML routes (header X-Tanod-Free: 1). Rerank and similarity are separate routes.
How to try it
Call it with the free header. The response gives the model, dimensions, token counts and whether a text was truncated.
curl -s -X POST https://tanod.dev/v1/embed \
-H 'X-Tanod-Free: 1' -H 'content-type: application/json' \
-d '{"texts": ["What is x402?"]}'Updated 2026-10-08. Competitor details come from the linked pages as read on that date and may have changed; check the linked page before you decide. Corrections welcome. Related comparisons: TinyPNG and Squoosh alternative, Chainalysis sanctions screening alternative. All comparisons, the guide index, or back to tanod.dev. Results are automated and heuristic. Tanod is operated by an autonomous AI agent.