Price and context comparisons
| Attribute | Nova Premier 1.0 | Command R7B |
|---|---|---|
| Input price | $2.50/M tokens | $0.0375/M tokens |
| Output price | $12.50/M tokens | $0.15/M tokens |
| Context window | 1,000,000 tokens | 128,000 tokens |
| Latency (p50) | Not available | 0.53 s |
| Throughput (p50) | Not available | 61.0 tok/s |
Command R7B costs $0.0375 per million input tokens and $0.15 per million output tokens on OpenRouter. Nova Premier 1.0 costs $2.50 per million input tokens and $12.50 per million output tokens. While Command R7B provides a lower price point for organizations processing massive amounts of text, Nova Premier 1.0 compensates with a context window that reaches 1,000,000 tokens to handle much larger datasets. Command R7B provides 128,000 tokens of context. Command R7B shows a latency of 0.53 seconds and a throughput of 61.0 tokens per second. Nova Premier 1.0 maintains a Multilingual Index of 79.6 percent.
Nova Micro is a text-only model that delivers the lowest latency responses. Nova Lite and Nova Pro provide 300,000 token context windows. Nova Premier provides 1,000,000 tokens of context for complex reasoning. Nova Micro, Nova Lite, and Nova Pro are at least 75 percent less expensive than the best performing models in their intelligence classes in Amazon Bedrock. Nova Lite performs equal or better than Anthropic Claude Haiku 3.5 on 10 of 12 benchmarks and performs equal or better than Google Gemini 1.5 Flash-8B on 17 of 21 benchmarks. Nova Pro matches or exceeds OpenAI GPT-4o on 17 of 20 benchmarks and matches or exceeds Google Gemini 1.5 Pro on 16 of 21 benchmarks.
Nova 2 and Forge customization
Amazon Web Services released the Nova 2 family in December 2025, comprising Nova 2 Lite, Nova 2 Pro, Nova 2 Sonic, and Nova 2 Omni. Nova 2 Lite is generally available and handles text, images, and videos for cost-conscious teams. It performs tasks like content summarization or video tagging in marketing pipelines. Nova 2 Pro remains in preview and tackles coding challenges or multi-step decision-making using text, images, videos, and speech. Nova 2 Sonic specializes in speech-to-speech dialogues, and Nova 2 Omni remains in preview to process text, images, video, and speech for text or image outputs. Nova Forge allows enterprises to build "Novellas" through pre-trained, mid-trained, or post-trained customization. This service costs $100,000 per year.
Nova 2 Lite and Nova 2 Pro use a one-million-token context window. Amazon Nova models integrate with Amazon Bedrock, which allows users to deploy models within a Virtual Private Cloud. Reddit, Sony, and Booking.com use these models for personalized recommendations, creative workflows, and travel suggestions. Nova 2 Lite provides reasoning focused performance. Nova 2 Pro acts as a reasoning agent for automating development cycles. Nova 2 Sonic handles customer service bots and virtual assistants, including accents and interruptions. Nova 2 Omni generates product visuals from voice descriptions.
You probably already know the differences between API-based and on-premise models. Nova Forge helps prevent "catastrophic forgetting" by using advanced techniques to preserve general intelligence while adding domain knowledge. Nova Forge uses training checkpoints to build custom models on Amazon SageMaker AI. This customization works across three stages: pre-trained customization for domain knowledge, mid-trained tweaks for balanced performance, and post-trained refinements for proprietary datasets. Amazon Nova Act provides general availability for building agents that take actions in a web browser.
Deployment and final verdict
Cohere offers on-premise and VPC deployment, which allows companies to keep data within their own infrastructure. Amazon Nova models require Amazon Bedrock for production access. Command R+ handles multi-step tool calls and provides in-line citations for verifiable responses. Command R+ also handles multi-hop retrieval across multiple documents. Command R features a tokenizer that compresses non-English text better than other models in the market. Cohere models serve business functions like finance, human resources, sales, marketing, and customer support. Command R and Command R+ are available in the US East (N. Virginia) and US West (Oregon) regions.
Command R7B wins for low-cost, high-throughput text tasks, while Nova Premier 1.0 wins for extremely long context needs. Which model provides the best return on investment for specialized business workflows? Command R+ provides a strong multilingual model for production-scale AI. Nova Premier 1.0 serves as the teacher model for distillation to create smaller, faster variants.




