All Services

AI / ML

Amazon Nova

Amazon Nova is Amazon's family of foundation models; current Nova 2 models are accessed through Amazon Bedrock for multimodal understanding, generation, and advanced reasoning workloads.

Explore pricing models, common use cases, infrastructure support, and the AWS services that commonly work with Amazon Nova.

Amazon Nova pricing and cost programs

Pricing model: Model, agent, and subscription usage

On-Demand
Available
Reserved Instances or reserved capacity
Service-specific
Savings Plans
Not applicable
Spot
Not applicable

Billing dimensions: Input and output tokens · Agent hours · Model customization · Subscriptions

Programs and modes: On-Demand model inference · Nova Act agent hours · Nova Forge subscription

Nova portfolio components use different token, agent-hour, and subscription pricing.

Free Tier: Service-specific — verify current offers

Pricing reviewed 2026-07-25. Reviewed against the linked official AWS pricing page. Recheck regional rates and program terms before purchase.

Official AWS pricing

Official AWS sources reviewed 2026-07-21.

Why implement Amazon Nova?

  • Offers Amazon-built models with different capability, latency, and cost profiles through the managed Bedrock runtime.
  • Supports multimodal inputs and model-dependent text, image, video, speech, or reasoning workflows without customer-hosted accelerators.
  • Uses Bedrock security, governance, retrieval, guardrail, evaluation, and observability integrations around model invocation.

How to implement Amazon Nova

  1. Choose the current Nova model that matches the modality, context, reasoning, latency, Region, and price requirements, then confirm model-specific request and output constraints.
  2. Build a versioned evaluation set that covers normal, edge, multilingual, adversarial, and safety cases; compare model and configuration candidates before committing.
  3. Invoke Nova through a least-privilege Bedrock workload, validate inputs and outputs, add guardrails and human escalation, and capture model ID, settings, latency, token use, and application outcomes.

Amazon Nova best practices

  • Follow the Nova responsible-use guidance, inform users where appropriate, and use qualified human review for outputs that affect rights, safety, finances, employment, health, or other consequential outcomes.
  • Do not infer reliability from fluent output; continuously test factuality, grounding, robustness, bias, and task-specific failure modes after model or prompt changes.
  • Minimize sensitive inputs, protect against prompt injection and unsafe tool execution, pin model versions where supported, and provide deterministic fallbacks for critical paths.

Amazon Nova use cases and server impact

  • Multimodal document and media understanding
  • Generative application features
  • Complex reasoning and agent workflows

Replaces self-hosted inference infrastructure for Nova models; the application, evaluations, safeguards, retrieval layer, tool permissions, and decision accountability remain yours.

Official implementation resources

Commonly paired AWS services