Cohere
Listed at https://dashboard.cohere.com
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Command R7: $2.50/M, Command R: $0.50/M, Embed: $0.10/M | per million tokens |
| Output | Command R7: $10/M, Command R: $1.50/M | per million tokens |
🤖 Supported Models (8)
✨ Pros
- ✓RAG pioneer, deep retrieval-generation integration
- ✓Command R7 with 128K context
- ✓Enterprise RAG solution (multi-step retrieval + tool use)
- ✓Embed + Rerank combo market leader for semantic search
- ✓Rerank model significantly improves search relevance
- ✓Excellent multilingual embedding performance
⚠️ Cons
- ×Command R7 pricing on the high side
- ×China requires proxy
- ×Limited model selection
- ×General chat capability behind GPT-4o
- ×Enterprise-focused, higher barrier for small teams
🎯 Best For
RAG system building; embeddings/reranking; enterprise search enhancement; multilingual search
💰 Pricing & Plans
| Model | Context | Input ($/M) | Output ($/M) | Notes |
|---|---|---|---|---|
| Command A+ (open-source) | 128K | Free | Free | Open-weight flagship, Apache 2.0, self-hostable |
| Command R7B | 128K | $0.0375 | $0.15 | Compact low-cost model for high-volume calls |
| Command R | 128K | $0.15 | $0.60 | RAG workhorse with citations & tool use |
| Command R+ | 128K | $2.50 | $10.00 | Legacy high-end, enterprise-grade |
| Embed 4 (text) | 128K | $0.12 | — | Semantic embeddings, 256-1536 dims |
| Rerank 4 Pro | 32K | $2.50 / 1K searches | — | Precision semantic ranking |
🔧 API & Developer Experience
- •API Style: Clean REST endpoints on dashboard.cohere.com with first-party Python and TypeScript SDKs, SSE streaming, and OpenAPI specs for codegen.
- •RAG-Native: Command models ship with connectors, grounded-generation endpoints and inline citations, so retrieval is a platform feature rather than a DIY layer.
- •Tool Use & Agents: Native function/tool calling on the Command family plus purpose-built Agent FMs (Command A Reasoning / Translate / Vision) for agent pipelines.
- •Embed + Rerank: The reference semantic-search pair: Embed 4 for recall, Rerank 4 for precision; used widely to upgrade existing vector DBs.
- •Free Trial: Signup auto-creates a rate-limited Trial API key (free, non-production); upgrade to pay-as-you-go for production.
- •Billing & Region: Billed in USD, monthly or on reaching a $250 outstanding balance; no free perpetual tier and no CNY settlement.
🎯 RAG & Enterprise Search (Command / Embed / Rerank)
Cohere's defining strength is enterprise retrieval-augmented generation (RAG). Rather than a general chatbot, the platform is built around the retrieval stack: Command models natively accept connected data sources, return grounded answers with inline citations, and drive multi-step tool use for production search and support automation. The Embed 4 and Rerank 4 models are the market's reference semantic-search pair — embeddings for recall and reranking for precision — which together lift retrieval quality far beyond pure vector search. Command A+ is also open-weight, so teams can self-host the flagship while keeping the managed APIs for embeddings and ranking. For teams whose real problem is finding the right answer across enterprise documents, Cohere's search-native pipeline is the most focused option on this list.
🌐 China Access & Latency
Cohere is a US-based, Western-first platform with no mainland-China point of presence, and its API is not directly reachable from within the country — the listing already flags 'proxy required.' Teams in mainland China must route through a stable overseas connection or a hosted relay, which adds latency and, depending on the route, can bring reliability issues during peak hours. Billing is in USD with monthly or $250-threshold invoicing, adding currency and tax friction compared with CNY providers. That said, once reachable from Hong Kong or Tier-1 corridors, the API is fast and reliable, and Cohere remains a reasonable choice for China-based teams whose priority is RAG and embeddings quality rather than local low-latency access. Treat it as an overseas service to be proxied, not a domestic endpoint.