Cohere API Pricing & Developer Hub
Everything you need to know about Cohere API pricing in 2026: platform fees, serverless quotas, compute instance rates, and bill optimization strategies.
Enterprise Intelligence: Grounded RAG, Embed v3 & Cross-Encoder Rerank
Cohere specializes in enterprise-grade language models designed specifically for factual retrieval, semantic search, and document reasoning. Unlike general consumer-focused model providers, Cohere's model family is optimized for corporate deployments where hallucination avoidance, multilingual understanding, and data residency are strict requirements.
The Cohere product suite consists of three foundational APIs: Command (Generation & RAG), Embed v3 (Dense Vector Retrieval), and Rerank v3.5 (State-of-the-art Cross-Encoder).
Cohere API Unit Economics (March 2026)
| Product / Model | Architecture Type | Input Rate / 1M | Output Rate / 1M | Search Unit Rate |
|---|---|---|---|---|
| Command R+ | 104B Frontier Enterprise RAG | $2.50 / 1M | $10.00 / 1M | N/A |
| Command R | 35B High-Efficiency RAG | $0.15 / 1M | $0.60 / 1M | N/A |
| Embed v3 (English / Multilingual) | 1024-dim dense embeddings | $0.10 / 1M | N/A | N/A |
| Rerank v3.5 | Cross-Encoder Relevance Scoring | N/A | N/A | $2.00 / 1k searches |
Frequently Asked Questions
Explore More Providers
Cohere Enterprise Intelligence: Embed v3, Rerank v3.5 & Command R+ Economics
Cohere is the enterprise pioneer in specialized search and retrieval architectures. While frontier LLM providers focus broadly on consumer chat interfaces, Cohere engineers high-precision enterprise information retrieval: **Embed v3** (dense multilingual vector embeddings with compression awareness), **Rerank v3.5** (state-of-the-art cross-encoder reranking at $1.00 per 1,000 searches), and **Command R+** ($2.50/M in, $10.00/M out) optimized for grounded tool-use and verifiable web citations across 128k context windows.
Cohere Production Model Rate Card
| Model ID | Primary Specialty | Input Rate | Output Rate | Context / Capabilities |
|---|---|---|---|---|
| embed-multilingual-v3.0 | Dense Embeddings | $0.10 / M tokens | N/A | 100+ languages / 1024 dim |
| rerank-v3.5 | Cross-Encoder Rerank | $1.00 / 1k queries | N/A | Up to 1,000 docs per query |
| command-r-plus-08-2024 | Enterprise Reasoning | $2.50 / M tokens | $10.00 / M tokens | 128k context / Citations |
| command-r-08-2024 | High-Speed RAG | $0.15 / M tokens | $0.60 / M tokens | Fast tool invocation |
Cohere Rerank v3.5 Client Implementation (Python)
import cohere
co = cohere.ClientV2(api_key=os.environ["COHERE_API_KEY"])
# Reorder top 25 candidate chunks by relevance, improving precision by 35%
results = co.rerank(
model="rerank-v3.5",
query="What is the data retention policy for EU healthcare tenants?",
documents=[doc["text"] for doc in candidate_chunks],
top_n=5
)
for hit in results.results:
print(f"Index: {hit.index} | Score: {hit.relevance_score:.4f}")Production SRE Guardrails
- Truncate Candidate Document Pools: Cohere bills per 1,000 search queries regardless of whether you rerank 5 documents or 100 documents. However, reranking more than 25 candidate chunks yields diminishing accuracy returns while increasing API response latency from 60ms to 240ms.
- Embedding Compression Types: Embed v3 supports native integer quantization (`int8` and `ubinary`). Requesting `embedding_types=['int8']` reduces vector database RAM footprint by 4x with zero loss in search accuracy.
Why is Cohere Rerank essential in production enterprise RAG?
Standard vector search retrieves chunks based on broad semantic proximity, frequently pulling irrelevant distractor documents. Cohere Rerank performs deep cross-attention between the query and each candidate, filtering out false positives and eliminating LLM hallucinations.
Can Cohere models be deployed on AWS Bedrock or Azure AI?
Yes. Cohere Command R+ and Embed models are natively available on AWS Bedrock, Microsoft Azure, and Oracle Cloud, allowing enterprise teams to bill usage against existing hyperscaler enterprise discount commitments.
Cohere Command-R+ Multi-Step Tool Calling & Enterprise RAG Grounding
Cohere's flagship Command-R and Command-R+ models are purpose-built for enterprise Retrieval-Augmented Generation (RAG) and multi-step agentic workflows. Unlike generic LLMs that hallucinate document citations, Command-R incorporates an intrinsic citation generation mechanism that outputs exact character span offsets corresponding to source context chunks. When paired with Cohere's Rerank 3.5 model, engineering pipelines first retrieve 100 candidate documents via dense vector similarity, pass them through Cohere Rerank to extract the top 5 semantically relevant passages, and feed only high-signal tokens into Command-R+, cutting downstream inference context token costs by up to 78%.
Cohere Private VPC Deployment & Sovereign Compliance Governance
For regulated financial services, healthcare institutions, and defense contractors requiring data sovereignty, Cohere offers complete VPC deployments across AWS Bedrock, Microsoft Azure AI, and Google Cloud Vertex AI. In these environments, inference requests never leave the enterprise's private virtual cloud perimeter, fully adhering to GDPR, HIPAA, and SOC 2 Type II strict audit requirements. Custom model fine-tuning with LoRA or full parameter updates can be executed against proprietary enterprise datasets without risk of weight leakage into public shared foundation models.
Frequently Asked Production & Architecture Questions
Q: How does Cohere Rerank 3.5 reduce downstream inference token costs?
A: By reranking an initial candidate pool of 100 documents retrieved via dense embeddings down to the top 5 semantically relevant passages, Rerank filters out noisy context and reduces input prompt token consumption by up to 78%.
Q: Can Cohere Command-R+ be deployed inside air-gapped enterprise virtual private clouds?
A: Yes. Cohere provides native private VPC deployments across AWS Bedrock, Microsoft Azure AI, and Google Cloud Vertex AI, ensuring inference requests and enterprise data never exit private virtual security boundaries.