Voyage AI gives every account the first 200 million tokens free on its current embedding models, and another 200 million free on its rerankers. No application, no card, no funding requirement. For a retrieval workload that is one of the largest free grants available anywhere, and it is oddly under-discussed next to the LLM signup credits everyone writes about.
Voyage builds the retrieval embedding and reranking models Anthropic recommends for RAG on Claude, which is the main reason it is worth knowing about if you are building on Claude specifically.
What is actually free
| Model family | Free allowance |
|---|---|
| voyage-4-large, voyage-4, voyage-4-lite, voyage-context-4, voyage-code-3 | First 200 million tokens per account |
| rerank-2.5, rerank-2.5-lite, rerank-2, rerank-2-lite | First 200 million tokens per account |
| voyage-multimodal-3.5, voyage-multimodal-3 | First 200M text tokens and 150 billion pixels |
| voyage-multilingual-2, voyage-finance-2, voyage-law-2, voyage-code-2 | First 50 million tokens |
Two things about that table are easy to misread, and both matter more than the headline number.
It is a one-time grant, not a monthly allowance. Unlike Groq's or Cerebras' daily free tiers, this does not reset. Once the 200 million are gone they are gone, which changes how you should spend them: on a real corpus, not trickled through six weeks of development.
The batch API does not draw it down — and that is a feature. Batch calls are billed at the discounted batch rate and do not consume the free balance. Which means the free grant and the cheap tier are complementary rather than sequential: use the free tokens for the latency-sensitive query path, and batch for the bulk corpus embedding you were going to do overnight anyway.
What 200 million tokens actually buys
Rough arithmetic, because the token count on its own means nothing:
- A typical prose document chunk runs around 500 tokens. 200 million tokens is therefore roughly 400,000 chunks.
- At 8 chunks per document, that is around 50,000 documents embedded — a substantial internal knowledge base, a full documentation site, or a mid-size support corpus.
- Query-side, embedding a search query is tens of tokens. Even at 100,000 queries a month you are spending a few million tokens a year on the query path.
So for most products the corpus is the expensive half and the queries are noise. If your corpus fits inside 200 million tokens, your embedding bill is plausibly zero for the life of the prototype, and you should be planning re-embedding runs (model upgrades, chunking changes) against the remaining balance rather than assuming an unlimited pool.
Where the rerankers matter more than the embeddings
The 200 million free reranker tokens are the more interesting half and get less attention. A reranker takes the 50 or so candidates your vector search returned and reorders them by actual relevance to the query. On most RAG systems, adding a reranker is a larger quality improvement than upgrading the embedding model, and it is the step teams skip because it adds a per-query cost.
With 200 million free reranker tokens, that objection goes away for the entire prototype phase. Reranking a query against 50 candidate chunks of 500 tokens costs roughly 25,000 tokens, so the free grant covers on the order of 8,000 reranked queries — enough to know whether it helps on your data before you pay for it.
How to claim it
- Create an account at dash.voyageai.com and generate an API key.
- Call any model in the voyage-4 family. The free tokens are consumed automatically — there is no coupon and no code to enter.
- Watch the usage meter in the dashboard. Because the allowance is per account and does not reset, the meter is the only thing standing between you and a surprise invoice.
- Move bulk corpus work to the batch API once the interactive path is working, so the free balance goes to the query path where latency matters.
We track the current terms on the Voyage AI perk page, re-checked against Voyage's own pricing documentation rather than a summary.
How it compares
Against the other free retrieval options:
- Cohere trial key — rate-limited trial rather than a token grant; good for evaluation, not for embedding a corpus.
- Google Gemini free tier — embeddings included in a broader free tier, with rate limits rather than a token ceiling.
- OpenAI embeddings — no free grant; you are paying from the first token, though text-embedding-3-small is cheap enough that this matters less than it sounds.
The wider comparison of what each embedding model costs per million tokens is in cheapest embeddings API, and the full free-tier retrieval stack — vector store, embeddings, reranker — is in RAG stack on free tiers.
For everything else that is free without an application, free AI API credits is the current list, each entry dated. Create a free account to see all 240 with their claim guides.