macrostack
Browse

The AI stack

The whole five-layer mapL3InfrastructureL4Models & toolingL5Applications

Categories

All 78 categories, by layer →

About

How we rank & score

Is OpenAI Embeddings API free? What it actually costs in 2026

The short answer

OpenAI embeddings are not free: text-embedding-3-small is $0.02 per million tokens and text-embedding-3-large $0.13, billed on input tokens with no free allowance.

The model pages list a price per million tokens and nothing else — no free tier, no monthly credit. Embedding a corpus is a one-off cost (a million tokens is roughly 750,000 words, so a small documentation set costs cents); re-embedding on every query or every re-index is where the meter runs.

What the free tier actually includes

  • No free tier or credits
  • text-embedding-3-small: $0.02 per 1M tokens
  • text-embedding-3-large: $0.13 per 1M tokens

Where the wall is

Re-indexing. The first pass over a corpus is cheap; the bill comes from embedding user queries at scale and from re-embedding the whole corpus whenever the chunking or model changes — which is also the moment a locally-run open model starts to pay for itself.

After that: text-embedding-3-small $0.02 per 1M tokens; text-embedding-3-large $0.13 per 1M tokens; usage is prepaid or billed monthly like the rest of the API.

Can you self-host OpenAI Embeddings API? No — cloud only

OpenAI's embedding models run only on OpenAI's servers. Nomic Embed and BGE-M3 are open-weight embedding models that run locally; Voyage AI is the hosted alternative compared here.

What it actually takes to run one yourself

Real RAM against the documented minimum, monthly cost, setup time, and the one thing that catches people out.

Self-hosting Nomic Embed

needs Around 1 GB of RAM or VRAM for one request at a time (estimate from the 274–547 MB weights; no figure is published). Batching long inputs raises it. (docs say Not stated by the project. The model is ~0.1B parameters — 547 MB of fp32 weights on Hugging Face, 274 MB as the F16 build in Ollama — and runs on CPU.)$0 — it runs on the machine you already own

Self-hosting BGE-M3

needs Roughly 3–4 GB of RAM on CPU, or 2–3 GB of VRAM with use_fp16=True, for one request at a time (estimate; no figure is published). Encoding at the full 8,192-token max_length in large batches needs far more, which is why the card tells you to lower max_length if you can. (docs say Not stated by the project. About 567M parameters: the pytorch_model.bin is 2.27 GB in fp32 and Ollama's build is 1.2 GB; it runs on CPU.)$0 — it runs on the machine you already own for a small corpus; a 24 GB RTX 3090/4090 on RunPod at $0.22–0.74/hr (read 2026-09-30) for a one-off bulk index

When paying is still the right call

Under roughly 10–15 million embeddings a month, the API is genuinely cheaper and simpler than self-hosting — $0.02/1M is a fair price and text-embedding-3-small is good. If your whole stack is OpenAI and your data isn't sensitive, staying is the pragmatic call.

What you are locked into

The sneaky kind: embeddings are model-specific, so switching means re-embedding the entire corpus into a new vector space. The API isn't holding your data hostage — the vectors are. Versioning your embedding model from day one is the escape hatch.

If you would rather not pay: Nomic Embed

92

Nomic Embed

The sovereign default — Apache-2.0, one command in Ollama, up to 8k context with rotary scaling.

Free — runs locally via Ollama or sentence-transformers; your hardware is the cost

See all 5 OpenAI Embeddings API alternatives compared

Compare the free options head-to-head

Nomic Embed vs BGE-M3→Nomic Embed vs Jina Embeddings v4→Nomic Embed vs Voyage AI→

Read from OpenAI's own pricing and documentation on 2026-09-27 — 3 days ago. Vendors change prices without notice; confirm before you commit.

The Macrostack brief

New swaps, worth your inbox.

A short, occasional email when we add a high-intent alternative or ship a new head-to-head. No spam, no selling your address — unsubscribe in one click.