Amazon Nova: AWS's New AI Models for Enterprise Bedrock Deployments

What You'll Learn

• What each Amazon Nova tier offers and how they compare on pricing
• How Nova pricing stacks up against Claude 3 Haiku and GPT-4o Mini
• Which business workloads belong on each Nova tier
• How to migrate existing Bedrock workflows to Nova in three steps
• The current limitations of Nova compared to more established models

Every executive responsible for cloud AI costs has felt the same pressure. The models that produce the best outputs are not the cheapest ones. The cheapest models require more prompt engineering and produce inconsistent results on complex tasks. The result: most organizations either overpay by routing everything to a premium model, or underpay and get unreliable outputs on important work.

Amazon's Nova model family arrives as a direct response to that problem. Three tiers: Nova Micro for text tasks at $0.035 per million input tokens, Nova Lite for multimodal tasks at $0.06, Nova Pro for complex reasoning at $0.80. At Nova Pro's pricing, Amazon delivers comparable performance to Claude 3 Haiku and GPT-4o Mini at 2 to 3 times lower cost.

For organizations already on AWS Bedrock, Nova is not a question of whether to evaluate. It is a question of how fast the evaluation can happen before next month's billing cycle.

SUBSCRIBER BREAK -- Premium Content Below

The Three Nova Tiers

Nova Micro ($0.035/million input tokens): Text-only. Optimized for classification, extraction of specific fields, yes/no routing decisions, and simple summarization of short text. Not the right choice for multi-step reasoning or creative generation.

Nova Lite ($0.06/million input tokens): Multimodal: accepts text, image, and video inputs. Handles document summarization, image analysis, chart and table extraction, and multi-modal data extraction from forms or invoices. The first model that makes multimodal processing economically viable at scale.

Nova Pro ($0.80/million input tokens): Highest accuracy in the Nova family. Still 2 to 3 times cheaper than Claude 3 Haiku or GPT-4o Mini at comparable performance. Handles complex multi-step reasoning, multi-document synthesis, code generation, and tasks where accuracy is critical.

Which Workloads Belong on Which Tier

Route to Nova Micro: document type classification, extraction of specific named fields, yes/no routing decisions based on clear criteria, entity recognition (company names, dates, dollar amounts). Route to Nova Lite: invoice processing including images, product catalog data extraction with images, form recognition from scanned documents, video content summarization. Route to Nova Pro: contract risk analysis, multi-document synthesis, code generation, financial reporting and compliance analysis where errors carry significant cost.

Migrating Existing Bedrock Workflows

• Identify your top workloads by token volume. Log into AWS Cost Explorer, filter by Bedrock usage. Identify the top three workflows by monthly token consumption. These are your migration priority.
• Map each workload to a Nova tier. Using the routing guide above, assign each to the appropriate tier. When uncertain between two tiers, start lower and test upward.
• Update the model identifier in your API calls. Nova Micro: amazon.nova-micro-v1:0. Nova Lite: amazon.nova-lite-v1:0. Nova Pro: amazon.nova-pro-v1:0. Run a parallel evaluation for 1 to 2 weeks before switching production traffic.

Current Limitations

Nova Micro supports up to 128K token context; Lite and Pro support up to 300K tokens. Nova is not a replacement for Gemini 1.5 Pro on very long document processing requiring 2-million-token windows. Nova's benchmarks are strong for a launch-day model, but expect some prompts that work on Claude or GPT-4o to need adjustment. At launch: available in us-east-1 only; organizations with strict data residency requirements should confirm region availability before migrating.