Cohere Command — Complete API Errors & Reference
Every Cohere API error, every model, every RAG-specific quirk — with the fix.
Current models & pricing (September 2026)
| Model | API ID | Input / M tok | Output / M tok | Context | Notes |
|---|---|---|---|---|---|
| Command R+ 2026 | command-r-plus-2026 | $3.00 | $15.00 | 256K tokens | Flagship generation model. Strong tool use, structured output, multilingual. Optimized for RAG grounding and citations. |
| Command R | command-r | $0.50 | $1.50 | 128K tokens | Mid-tier. Best price/performance for RAG. Extremely cost-effective for high-volume workloads. |
| Command Light | command-light | $0.15 | $0.60 | 64K tokens | Fast and cheap. Classification, simple extraction, high-throughput routing. |
| Aya Expanse 32B | c4ai-aya-expanse-32b | $0.50 | $1.50 | 128K tokens | Multilingual specialist. Strong performance across 23 languages including Arabic, Hindi, Chinese, Turkish. |
| Embed v4 | embed-english-v4.0 | $0.10 | n/a | 8K tokens | Embedding model. 1024-dim vectors. Best-in-class for retrieval on standard benchmarks. |
| Rerank 4 | rerank-v4.0 | $2.00/1K docs | n/a | 4K tokens/doc | Reranking model. Reorders retrieval results by relevance. Standard 2nd-stage RAG component. |
Rate limit tiers
| Tier | Requirement | Limits |
|---|---|---|
| Trial | Free account, no billing | 20 RPM (chat) · 100 RPM (embed) · limited monthly quota |
| Production | Billing enabled | 10,000 RPM (chat) · 2,000 RPM (embed) · higher quotas |
| Enterprise | Contact sales | Custom RPM · dedicated capacity · private deployment · SLA |
Documented errors
Related pillar guides
Related blog posts
Frequently asked questions
When should I choose Cohere over Anthropic or OpenAI?
Two main cases: (1) RAG workloads where Embed v4 + Rerank 4 is a proven component stack — many production systems use these regardless of which LLM they generate answers with, (2) high-volume production where Command R's $0.50/$1.50 pricing dramatically undercuts equivalent tiers from American flagship providers. For reasoning-heavy workloads, Anthropic and OpenAI are still typically stronger.
What's Cohere's differentiation in 2026?
Enterprise RAG excellence and multilingual coverage. Embed v4 leads standard retrieval benchmarks; Rerank 4 is widely adopted as the reranking component in production stacks; Aya Expanse handles 23 languages well. Cohere positions as the RAG-first provider, and their pricing reflects that focus.
What model should I use for a new Cohere integration?
Command R for most generation workloads — strong performance at $0.50/$1.50. Command R+ 2026 when you need flagship-tier reasoning and tool use. Embed v4 for embeddings (regardless of which generation model you use). Rerank 4 for retrieval reranking (universal component).
Can I use Cohere Embed and Rerank with a different LLM?
Yes, and many teams do exactly this. Common pattern: Cohere Embed v4 for indexing → vector search → Cohere Rerank 4 for top-K reordering → Claude Sonnet 4.6 or GPT-5.6 for generation. Cohere is often the retrieval layer under an American flagship generation layer.
What's the difference between Cohere's v1 and v2 APIs?
v2 (2024+) is the current recommended API with cleaner semantics, better tool use, and improved citations. v1 endpoints are deprecated with published sunset dates. Migrate to v2 when you refactor RAG code — v2 handles multi-turn tool use much more cleanly.
How does Cohere handle citations for RAG?
Native citations output — when you pass documents to the Chat endpoint, responses include a citations array with document IDs and character spans in the response. This grounds responses in retrieval sources without extra prompt engineering. See docs.cohere.com/reference/chat for the citations schema.
Does Cohere offer private deployment?
Yes — enterprise tier includes options for deployment on your cloud (AWS, Azure, GCP) with dedicated capacity and full data isolation. This is a major differentiator vs API-only providers for regulated industries. Contact Cohere sales for the current deployment models.
How do Cohere's rate limits differ across endpoints?
Chat, Embed, and Rerank have separate rate limits. Production tier: 10,000 RPM for Chat, 2,000 RPM for Embed, ~1,000 RPM for Rerank. When building RAG stacks, embed calls typically dominate — plan quotas accordingly.