Amazon Nova

Definition

Amazon Nova is the family of foundation models created by AWS and released in late 2024. The Nova family covers text, multimodal, image, and video generation, and is available exclusively on Amazon Bedrock. Nova is positioned as the price/performance leader for high-volume tasks where frontier-model reasoning is not required.

The Nova family

Nova Micro — Text-only model optimized for ultra-low latency and lowest cost. Per-million input tokens ≈ $0.04. Use for high-volume classification, routing, and short summarization.

Nova Lite — Multimodal (text + images + video). Strong price/performance for OCR, document understanding, video summarization. Per-million input tokens ≈ $0.06.

Nova Pro — Higher reasoning capability with multimodal input. Suitable for complex extraction, multi-step reasoning, and tool use. Per-million input tokens ≈ $0.80.

Nova Premier — Most capable Nova reasoning model, for tasks that demand frontier-level reasoning at lower cost than Claude Opus.

Nova Canvas — Image generation, editing, and inpainting. Includes invisible watermarking for AI provenance.

Nova Reel — Short-form video generation (6 seconds at GA, extending in 2025–2026 releases). Includes invisible watermarking.

When to use Nova

• High-volume classification, routing, extraction — Nova Micro beats Claude Haiku on cost for these workloads.
• Document and video understanding at scale — Nova Lite is a strong default.
• Cost-sensitive multi-tenant SaaS GenAI — Nova lets you offer GenAI features within tight unit-economic budgets.

When not to use Nova

• Complex reasoning over long documents — Claude Sonnet 4 / Opus 4 typically outperform Nova Pro on benchmarks like LongBench and MMLU-Pro.
• Fine-tuning for proprietary domains — Llama 4 is the open-weight default; Nova does not currently support customer fine-tuning.
• Tool-use-heavy agents — Claude with native tool use remains the safer default.

Common mistakes

Mistake 1: Picking Nova Pro or Premier without benchmarking against Claude. Nova is often the right call on cost, but you owe yourself a 100-example evaluation against the use case.

Mistake 2: Using Nova Reel for long-form video. The model targets short clips. For long-form, plan multi-clip generation with a separate stitching pipeline.

Mistake 3: Treating Nova as a single model. The price gap between Micro and Premier is ~150×. Pick the smallest Nova that meets quality bars.

Related AWS Services

• Amazon Bedrock — Hosts all Nova models
• Bedrock Prompt Caching — Works with Nova for cost reduction on large system prompts
• Bedrock Guardrails — Apply consistent safety policy across Nova and non-Nova models

Related FactualMinds Content

• Amazon Bedrock Consulting
• Generative AI on AWS