No — Anthropic does not offer its own embedding model. Anthropic’s own documentation states it plainly: “Anthropic does not offer its own embedding model,” and points developers to Voyage AI as its recommended embeddings partner instead. (Checked 2026-09-08 via platform.claude.com/docs/en/build-with-claude/embeddings.) That’s the accurate answer, but it’s not the only option. The table below compares Voyage AI against two open-source embedding models you can call through an OpenAI-compatible endpoint on Novita AI, without adding a second SDK to your stack.
| Model | Price per million tokens | Context length | OpenAI-compatible endpoint |
|---|---|---|---|
Voyage voyage-4 (Anthropic’s recommended vendor) | $0.06 | 32,000 tokens | No — uses Voyage’s own SDK/REST format |
| BAAI BGE-M3 (Novita AI) | $0.01 | 8,192 tokens | Yes — /openai/v1/embeddings |
| Qwen3 Embedding 8B (Novita AI) | $0.07 | 32,768 tokens | Yes — /openai/v1/embeddings |
(Checked 2026-09-08. Voyage voyage-4 pricing and context length via docs.voyageai.com/docs/pricing and platform.claude.com/docs/en/build-with-claude/embeddings. BGE-M3 and Qwen3 Embedding 8B pricing/context via novita.ai/pricing: BGE-M3 listed at context 8192, input pricing $0.01/Mt; Qwen3 Embedding 8B listed at context_size 32768, input pricing $0.07/Mt. Endpoint compatibility via raw.githubusercontent.com/novitalabs/novita-skills/main/skills/novita-ai/references/llm-api.md: POST /openai/v1/embeddings.)
Why Anthropic Doesn’t Ship an Embedding Model
Claude is built for generation and reasoning — chat, tool use, agentic workflows. Embeddings are a different problem: turning text into a fixed-length vector for similarity search, clustering, or retrieval. Rather than building and maintaining a separate embedding model, Anthropic’s docs recommend assessing “a variety of embeddings vendors to find the best fit for your specific use case,” with Voyage AI named as the primary option covered in their guide.
This matters for how you architect a Claude-based RAG pipeline: you’ll always need a second API call to a non-Anthropic endpoint for the embedding step, whether that’s Voyage or something else.
Option 1: Voyage AI (Anthropic’s Recommended Path)
| Model | Context length | Price per million tokens | Free tier |
|---|---|---|---|
voyage-4-large | 32,000 tokens | $0.12 | First 200M tokens free |
voyage-4 | 32,000 tokens | $0.06 | First 200M tokens free |
voyage-4-lite | 32,000 tokens | $0.02 | First 200M tokens free |
(Checked 2026-09-08 via platform.claude.com/docs/en/build-with-claude/embeddings for model names/context length, and docs.voyageai.com/docs/pricing for the free-tier and per-token figures: “The first 200 million tokens for voyage-4-large, voyage-4, voyage-4-lite… are free for every account.”)
Voyage AI’s endpoint is POST https://api.voyageai.com/v1/embeddings, using its own SDK or a direct REST call — not the OpenAI SDK format. It’s a solid choice if you want Anthropic’s endorsed vendor and don’t mind a separate account and billing relationship for embeddings.
Option 2: OpenAI-Compatible Open-Source Embeddings on Novita AI
If you’d rather keep everything behind one OpenAI-compatible client instead of adding a Voyage-specific SDK, Novita AI hosts open-source embedding models behind the same /openai/v1/embeddings endpoint you’d already use for chat completions.
BGE-M3 supports over 100 languages and combines dense, multi-vector, and sparse retrieval in one model, with an 8,192-token context window — enough for most document chunks in a RAG pipeline. (Checked 2026-09-08 via novita.ai/pricing: “Capable of processing over 100 languages and inputs ranging from short sentences to lengthy documents (up to 8,192 tokens)… topping benchmarks like MIRACL and MKQA.”)
Qwen3 Embedding 8B ranked No. 1 on the MTEB multilingual leaderboard with a score of 70.58 as of its June 2025 release, and supports a longer 32,768-token context window than BGE-M3. (Checked 2026-09-08 via Hugging Face’s Qwen/Qwen3-Embedding-8B model card, which reports the 70.58 MTEB multilingual score and No. 1 ranking.) It costs more per token than BGE-M3 but handles longer inputs without chunking.
Calling It From Code That Already Uses OpenAI’s SDK
from openai import OpenAI
import os
client = OpenAI(
base_url="https://api.novita.ai/openai/v1",
api_key=os.environ["NOVITA_API_KEY"],
)
response = client.embeddings.create(
model="baai/bge-m3",
input="Anthropic recommends Voyage AI for embeddings, but this text is being embedded with BGE-M3 instead.",
)
print(response.data[0].embedding[:5]) # first 5 dimensionsSwap model="baai/bge-m3" for model="qwen/qwen3-embedding-8b" to use the longer-context model. Both accept a string or array of strings as input and return data[].embedding as a float array, matching the OpenAI embeddings response shape. (Checked 2026-09-08 via raw.githubusercontent.com/novitalabs/novita-skills/main/skills/novita-ai/references/llm-api.md: POST /openai/v1/embeddings, parameters input and model, “Returns data[].embedding (float array).”)
Which One Fits Your Pipeline
If you’re already deep into Anthropic’s ecosystem and want the vendor they explicitly vet, Voyage AI is the path with the least friction — you get Matryoshka-style variable-dimension output and models tuned for code and multilingual retrieval. If your pipeline is OpenAI-SDK-based (LangChain, LlamaIndex, or a custom client pointed at an /embeddings endpoint) and you want to avoid a second SDK just for the embedding step, BGE-M3 or Qwen3 Embedding 8B on Novita AI slot in with a base URL change and no new client library.
Cost is the other lever: BGE-M3 at $0.01 per million tokens is cheaper than every Voyage model, including voyage-4-lite at $0.02. If your context windows fit inside BGE-M3’s 8,192-token limit, that’s the lowest per-token cost of the four options compared here. If you need to embed longer chunks without splitting them, Qwen3 Embedding 8B’s 32,768-token window matches Voyage’s voyage-4 family at a similar per-token price ($0.07 versus $0.06).
FAQ
Does Claude have a built-in embeddings feature? No. Anthropic’s documentation states directly that Claude does not offer an embedding model and directs developers to third-party providers, primarily Voyage AI.
Can I use Voyage AI embeddings with Claude in the same pipeline? Yes — that’s the intended pattern. You call Claude for generation and Voyage AI separately for embeddings; there’s no combined endpoint.
Is there a cheaper alternative to Voyage AI that still works with OpenAI-style code?
Yes. Open-source models like BAAI BGE-M3 and Qwen3 Embedding 8B are hosted on Novita AI behind an OpenAI-compatible /embeddings endpoint, at $0.01 and $0.07 per million tokens respectively, both below Voyage’s mid-tier pricing.
Which context window should I pick? BGE-M3 caps at 8,192 tokens per input; Qwen3 Embedding 8B and Voyage’s current-generation models both support 32,000+ tokens. Pick based on your largest chunk size, not your average one.
