Command R+ on Oracle Cloud Infrastructure provides the most direct path for enterprises that require high-volume retrieval-augmented generation and private deployment. The partnership allows users to run models on OCI cluster GPUs to manage workloads at high performance and low cost. Oracle’s vector database allows for better semantic search without moving data.
OCI Enterprise AI development steps
OCI Enterprise AI follows a four-step workflow for developers. Teams first choose the foundation model that fits their use case and performance needs. They then connect the agent to enterprise data and knowledge sources, including structured systems and vector search for unstructured content. Developers define the agent workflow by selecting the tools and APIs the agent calls to orchestrate multi-step actions and conversation context. Finally, teams deploy and operate the agent in production with identity and access management, guardrails, observability, and auditability.
The integration of Cohere models into Oracle Cloud Infrastructure provides a complete, end-to-end platform for generative AI. This platform uses advanced security and best-in-class data management. Users can leverage OCI’s cluster GPU to run AI workloads at the highest performance.
Deployment on dedicated AI clusters
New dedicated AI clusters require specific hardware unit shapes. To reach a model through a dedicated AI cluster, users must create an endpoint for that model. For the Command R+ 08-2024 model, the OCI model name is cohere.command-r-plus-08-2024. This model uses the Large Cohere V2_2 unit size and requires one unit. The pricing for this configuration is listed as Large Cohere – Dedicated. For the Command R+ 08-2024 (tp4) variant in the UAE East (Dubai) region, the OCI model name is cohere.command-r-plus-08-2024-tp4-64k. This version uses the Large Cohere unit size and also requires one unit. The hosting cost for this Dubai variant multiplies the unit price by two. The limit name is dedicated-unit-large-cohere-count, and the request limit increase is two.
The Command A family also requires specific configurations. The model name Cohere Command A uses the OCI model name cohere.command-a-03-2025. This model uses the LARGE_COHERE_V3 unit size and requires one unit. For the Command A variant in the UAE East (Dubai) region, the model uses the SMALL_COHERE_4 unit size and requires one unit. In the Dubai region, the hosting price multiplies by four and the request limit increase is four. The limit name for the Command A variant is dedicated-unit-small-cohere-count.
For legacy generic Cohere shapes being retired, these shapes remain available in the API only. Users with a dedicated AI cluster using a legacy generic Cohere shape may see both the legacy shapes and the new hardware unit shapes until the service removes the legacy shapes. For new dedicated AI clusters, users must use the hardware unit shapes listed in the regional documentation.
The retrieval-augmented generation pipeline
Cohere provides a vertically integrated RAG stack consisting of Embed, Rerank, and Command. Embed converts documents and queries into vector representations. Rerank re-scores retrieved results for relevance before the generation model receives them. Command R+ acts as the generation layer by receiving the reranked context. Command R+ features a 128,000 token context window. This architecture distinguishes between retrieved context and parametric knowledge to reduce hallucination risk. The model produces traceable, structured citations without explicit prompting. You know the basics of RAG, so focus on the deployment architecture.
A cheap generation model can still produce an expensive RAG system if every request retrieves too many chunks, reranks a large candidate set, and sends duplicated evidence into the prompt. To avoid this, developers should limit candidate counts and avoid unnecessary replicas. Index refreshes also impact costs, as re-embedding an entire corpus after small edits wastes API calls and vector database writes. Using stable document identifiers and content hashes helps prevent these costs. Rerank quality depends on the quality of the initial candidates, so a reranker cannot recover documents that retrieval never returned.
| Model | Context Window | Input Cost (per 1M) | Output Cost (per 1M) |
|---|---|---|---|
| Command R | 128,000 | $0.50 | $1.50 |
| Command R+ | 128,000 | $2.50 | $10.00 |
| Command A | 256,000 | $2.50 | $10.00 |
| Claude 3.5 Sonnet | 200,000 | $3.00 | $15.00 |
Multimodal extraction with Parse 5
Parse 5 released on August 27, 2026, serves as a vision language model with 2.3 billion parameters. It converts visually rich PDFs, PPTs, and JPEGs into Markdown. The model uses an 8,192 token context window. It utilizes a 400M parameter native-resolution vision encoder initialized from SigLIP 2 SO400M. This encoder uses 2D Rotary Positional Embeddings and learned 1D positional embeddings to preserve spatial document structure. A dedicated Projector maps extracted visual features into the language model embedding space. The reasoning engine uses an in-house 2B parameter language model based on Command A+ architecture.
The Parse 5 model achieves a 79.2 score on the ParseBench benchmark by successfully managing table extraction, content faithfulness, and semantic formatting across thousands of human-verified enterprise pages. The API accepts image URLs or data URIs. Current versions do not support PDF or file URLs directly through the POST /v2/parse endpoint. Files must be 20 MB or less, and the decoded size cannot exceed 50 megapixels or 200 MB. Users can access Parse 5 via Cohere’s platform, Microsoft Azure AI Foundry, or Amazon SageMaker. Model Vault Parse 5 instances cost $4.00 per hour for Medium or $7.00 per hour for XL. Monthly rates for these instances are $2,500 and $4,300.
Comparison of Command R+ and Claude
Command R+ is purpose-built for retrieval-augmented generation. It produces structured citations by default, whereas Claude requires explicit prompt engineering to cite sources accurately. Claude uses instruction-based grounding, which relies on system prompts to control hallucinations. Command R+ features a 128,000 token context window, while Claude provides a 200,000 token context window. Claude excels at general reasoning, coding, and complex multi-step instruction following.
Command R+ provides a structural advantage through private deployment. It is available as a model weight download for on-premises or private cloud environments. This allows data to stay within organizational infrastructure, which satisfies requirements for healthcare, legal, and financial industries. Anthropic does not offer private deployment for Claude. Claude requires the Anthropic API, AWS Bedrock, or Google Cloud Vertex AI. The comparison shifts on AWS Bedrock, as both models run within AWS infrastructure, removing the private deployment advantage.
| Model | Input $/M | Output $/M | Context |
|---|---|---|---|
| Command R+ | $2.50 | $10.00 | 128,000 |
| Command R | $0.50 | $1.50 | 128,000 |
| Claude 3.5 Sonnet | $3.00 | $15.00 | 200,000 |
| Claude 3 Opus | $15.00 | $75.00 | 200,000 |
Implementation and integration workflows
Implementation requires planning for throughput and infrastructure. For the HumanFirst integration, users first obtain a Cohere API key from the Cohere website. They then create a new namespace in HumanFirst Studio and add the Cohere integration. Users set a name and a rate limit for the integration. The free plan allows 100 calls per minute with 16 embeds per call. Users select a model and then create a workspace.
Enterprise teams building on the OCI platform can deploy Cohere Command R+ using dedicated clusters to ensure that their inference traffic stays within their own private VPC or on-premises environment for regulatory compliance. Companies can also use Cohere’s managed API, dedicated Model Vault, or cloud marketplaces like AWS Bedrock. For those using AWS Bedrock, Bedrock Rerank uses the Cohere Rerank model.
| Feature | Cohere Command R+ | Anthropic Claude |
|---|---|---|
| Deployment | Managed, VPC, On-prem | Managed API, Bedrock, Vertex AI |
| RAG Focus | Native grounding/citation | Prompt-based grounding |
| Context Window | 128,000 tokens | 200,000 tokens |
| Primary Use | Retrieval-heavy tasks | Reasoning and coding |
Industry applications and scaling
Cohere’s models support various enterprise workflows. In asset management, users use Command and Rerank to create virtual assistants that answer analyst questions from cloud repositories. Retailers use Embed to automate the collection of data from support tickets and user reviews in over 100 languages. Staff support applications use Command and Rerank to integrate with HR databases for employee queries. Online shopping assistants use Command and Rerank to manage product catalogs and customer service applications.
The total cost of ownership involves more than just token rates. Developers must account for deployment effort, implementation cost drivers, and support exposure. Private deployment requires customer-procured GPU hardware, Kubernetes or equivalent orchestration, and ongoing operations ownership. For high-volume workloads, Command R pricing is optimized to reduce costs. For specialized tasks, Command R7B provides an economical option for narrow work. Which specific regional availability will Cohere announce next for the Command A family?
