Cohere (Toronto, founded 2019 by "Attention Is All You Need" co-author Aidan Gomez and ex-Google Brain researchers) built its platform around one use case: enterprise retrieval and RAG. Most developers don't realize the three things Cohere leads on are all free on a trial key.

• Embed v3 — top-of-MTEB text embeddings, English and 100+ languages
• Rerank v3 — the most-deployed neural reranker in production RAG, one API call
• Command R / R+ — chat models trained for RAG, tool use, and grounded citations

What's Free: Trial vs Production Keys

Trial keys are free and never expire, but they are capped at 1,000 API calls a month on top of the per-minute rate limits below. Cohere's rate-limit docs state this plainly, and it is the number that decides whether the free tier fits your project — not the per-minute figures most write-ups quote. (Verified 2026-09-01.)

Embed is metered at 2,000 inputs/min, but the monthly cap binds first: at up to 96 documents per call, 1,000 calls is roughly 96,000 embeddings a month. Enough to index a personal knowledge base once; not enough to re-index it nightly. The old /v1/classify and /v1/summarize endpoints were retired on 2025-09-15 — use Chat for both.

Get Your Free Key

1. Sign up at dashboard.cohere.com/welcome/register (email or Google)
2. Verify your email, then open API Keys in the sidebar — your default Trial key is already there
3. export COHERE_API_KEY="your_key_here"

No credit card, no phone number.

Python Quickstart: First Embedding

pip install cohere

import os
import cohere

co = cohere.ClientV2(api_key=os.environ["COHERE_API_KEY"])

response = co.embed(
    texts=[
        "Cohere makes the best free embedding API for RAG.",
        "Toronto is the headquarters of Cohere."
    ],
    model="embed-english-v3.0",
    input_type="search_document",
    embedding_types=["float"]
)

print(f"Got {len(response.embeddings.float)} embeddings")
print(f"Each is {len(response.embeddings.float[0])} dimensions")

That returns 1024-dimensional vectors for any vector DB (Pinecone, Weaviate, Chroma, Qdrant, pgvector). Key detail: Cohere embeddings are asymmetric — use input_type="search_document" when indexing and "search_query" when embedding the user's question. This beats symmetric APIs on retrieval quality.

Embedding Models (all free)

embed-english-v3.0 (1024d) is the sweet spot. The light variants drop to 384d (~60% smaller index) with only a small quality loss.

Rerank: Cohere's Secret Weapon

After your vector DB returns the top 50–100 candidates, pass them to Rerank with the query. It scores true relevance and reorders — the top few reranked results are usually far better than raw vector similarity.

co = cohere.ClientV2(api_key=os.environ["COHERE_API_KEY"])

response = co.rerank(
    model="rerank-english-v3.0",
    query="How do I add a free embedding API to my chatbot?",
    documents=[
        "Cohere offers free embedding API access through trial keys.",
        "Pinecone is a managed vector database service.",
        "Use embed-english-v3.0 for the best quality English embeddings.",
    ],
    top_n=3
)

for r in response.results:
    print(f"Score: {r.relevance_score:.4f}  |  index {r.index}")

Adding a Rerank step typically boosts RAG answer quality by 15–30% over vector-similarity-only retrieval. Free on the trial key: 10 calls/min, up to 1,000 documents per call.

Chat with Command R+: Built for Citations

Command R+ accepts a structured documents parameter and returns inline citations pointing to which document each claim came from — far more useful than stuffing docs into a system prompt for legal, medical, or internal-KB use.

response = co.chat(
    model="command-a-03-2025",
    messages=[{"role": "user", "content": "Which Cohere embedding model for English RAG?"}],
    documents=[
        {"data": {"text": "embed-english-v3.0 is 1024-dim and leads MTEB English."}},
        {"data": {"text": "embed-english-light-v3.0 is 384-dim, low storage cost."}},
    ]
)

print(response.message.content[0].text)
for c in response.message.citations or []:
    print(f"  - '{c.text}' from {[s.id for s in c.sources]}")

All three run on the free key at 20 req/min. Note the model IDs: Cohere deprecated the bare command-r and command-r-plus aliases on 2025-09-15, so pin a dated ID or command-a-03-2025, which Cohere calls its strongest model across domains. (Verified 2026-09-01.)

End-to-End RAG Pipeline (All Free)

Embed, store, retrieve, rerank, and answer with citations — on one trial key, no credit card:

import os, numpy as np, cohere
co = cohere.ClientV2(api_key=os.environ["COHERE_API_KEY"])

documents = [
    "Cohere Embed v3 produces 1024-dim vectors optimized for retrieval.",
    "Cohere Rerank v3 reorders candidates by true relevance to the query.",
    "Command R+ is a 104B model trained for RAG with citations.",
    "Cohere trial keys never expire but are capped at 1,000 API calls a month.",
]

# Index
doc_matrix = np.array(co.embed(
    texts=documents, model="embed-english-v3.0",
    input_type="search_document", embedding_types=["float"]
).embeddings.float)

# Embed query + vector similarity for top 3
query = "How do I get free access to Cohere's RAG models?"
q = np.array(co.embed(texts=[query], model="embed-english-v3.0",
    input_type="search_query", embedding_types=["float"]).embeddings.float[0])
top = np.argsort(doc_matrix @ q)[-3:][::-1]
candidates = [documents[i] for i in top]

# Rerank to best 2, then answer with grounded citations
reranked = co.rerank(model="rerank-english-v3.0", query=query,
    documents=candidates, top_n=2)
top_docs = [candidates[r.index] for r in reranked.results]

answer = co.chat(model="command-a-03-2025",
    messages=[{"role": "user", "content": query}],
    documents=[{"data": {"text": d}} for d in top_docs])
print(answer.message.content[0].text)

Tip: with only a few hundred candidate docs, skip the vector-DB step and pass everything straight to Rerank (up to 1,000 docs/call).

JavaScript / Node.js

npm install cohere-ai

import { CohereClientV2 } from "cohere-ai";
const co = new CohereClientV2({ token: process.env.COHERE_API_KEY });

const response = await co.embed({
  texts: ["Cohere is the best free embedding API for RAG."],
  model: "embed-english-v3.0",
  inputType: "search_document",
  embeddingTypes: ["float"]
});
console.log(`Got ${response.embeddings.float.length} embeddings`);

Cohere vs Other Free Embedding Options

Cohere is the only free tier here that ships a hosted neural reranker — which, combined with asymmetric embeddings, usually matters more for RAG quality than the base embedding model.

Pricing (When You Outgrow Free)

Command R7B and Embed v3 are among the cheapest production-grade options in their class.

When to Use Cohere (and When Not To)

Use it when: you're building RAG and want the best free embeddings + reranker combo; you need multilingual retrieval across 100+ languages; you require grounded citations (legal, medical, internal KBs); or you want asymmetric query/document embeddings.

Look elsewhere when: you need raw chat throughput (Groq for free speed, Gemini Flash for free quota); OpenAI SDK drop-in compatibility (Mistral, DeepSeek); or image/audio/multimodal generation (Cohere is text-only).

Frequently Asked Questions

Does the free Cohere key expire?

No — trial keys never expire and need no credit card. They are capped at 1,000 API calls a month plus per-minute rate limits, which makes them right for prototyping and side projects but not for production traffic.

What's the catch with the free tier?

Two limits, and the monthly one is the one that bites: 1,000 API calls a month, plus per-minute rate limits (20 req/min chat, 10 req/min rerank; Embed is metered at 2,000 inputs/min). Trial keys are meant for building and evaluating, not for production traffic.

Why use Cohere over OpenAI or Gemini for RAG?

Cohere is the only major free tier that includes a hosted neural reranker, plus asymmetric embeddings and a chat model that returns inline citations. That covers more of the RAG pipeline than any other single free provider.

What is the input_type parameter for?

Cohere embeddings are asymmetric. Use search_document when indexing your corpus and search_query when embedding the user's question — this noticeably improves retrieval quality over symmetric embedding APIs.

Final Verdict

Cohere is the most underrated free AI API because it ships a complete RAG stack — embeddings, reranker, and a citation-trained chat model — all behind one free trial key. Most "free API" roundups skip it by comparing only chat models, missing what the company actually built. If your project involves search over your own documents, Cohere's free tier covers more of the pipeline than any other single provider. Sign up at dashboard.cohere.com and your first reranked retrieval is about ten minutes away.