Amazon Nova

Definition

Amazon Nova is AWS’s first-party foundation model family, available exclusively through Amazon Bedrock. The current everyday Nova is Nova 2 Lite (GA December 2, 2025): multimodal text, image, and video with a 1M-token context window (amazon.nova-2-lite-v1:0). Nova 2 Sonic covers speech. Nova 2 Pro is preview — do not treat it as general GA. Prior-generation text and multimodal models remain on the catalog: Nova Micro (lowest latency and published cost), Nova Lite, Nova Pro, and Nova Premier. Nova Canvas generates and edits images; Nova Reel generates short-form video — see Nova Canvas and Reel guide. Nova is positioned for high-volume workloads where frontier-model reasoning from Claude Sonnet 5 or Opus 5 is unnecessary — classification, routing, OCR-heavy extraction, and cost-sensitive multi-tenant features.

When to use it

• Everyday multimodal at volume — customer support, document processing, and business automation on Nova 2 Lite.
• Cheapest text-only routing — intent classification and pre-screens on Nova Micro when the 1M window is unused.
• Document and media understanding at scale — PDFs, slides, charts, and short video (Nova 2 Lite or prior Nova Lite).
• Cost-sensitive SaaS features where unit economics require a lower per-token model than Claude Sonnet 5.5 or Haiku 5.5 (Haiku 4.5 remains the prior pin; benchmark Nova 2 Lite vs Nova Pro / Premier).
• Image generation and editing in AWS-native workflows without third-party image APIs (Nova Canvas).
• Short marketing or training clips where Nova Reel’s clip length fits the use case.
• Speech in/out — Nova 2 Sonic, not Nova 2 Lite.

When not to use it

• Long-horizon agentic workflows with heavy tool use — Claude Sonnet 5.5 or Opus 5.5 remain safer defaults until you validate Nova on your tool-calling traces. Fable 5.1 is a Covered Model path, not a Nova substitute.
• Assuming Nova 2 Pro is GA — it is preview; do not pin production on preview access.
• Customer fine-tuning of proprietary weights — Nova does not offer the same open-weight fine-tuning path as Llama on Bedrock; use customization features only where documented for Nova.
• Long-form video production — Nova Reel targets short clips; multi-minute video needs external pipelines or other services.

Tips

• Benchmark Nova 2 Lite vs Nova Micro on a representative sample before defaulting to 2 Lite for text-only routing — you still pay for unused context.
• Pair Nova with Bedrock Guardrails and Prompt Caching for repeated system prompts in multi-tenant apps.
• Use Nova Premier only after Sonnet 5.5 / Haiku 5.5 (Haiku 4.5 remains the prior pin) / Nova 2 Lite comparisons on your eval set.
• For Canvas and Reel, plan content moderation and watermarking requirements upfront — Nova includes provenance features AWS documents for generated media.
• Log model ID and token usage per tenant for chargeback and abuse detection.
• Verify Nova 2 list rates on Bedrock pricing — do not copy unpublished $/MTok from blogs.

Gotchas

Serious

• Assuming Nova Pro or Nova 2 Lite matches Opus 5 on complex reasoning — run task-specific evals; legal, medical, and multi-hop analysis often still favor Claude.
• Feeding unredacted PHI into the wrong environment — Nova follows Bedrock data handling; your compliance boundary still requires BAA, logging, and access controls.

Regular

• Using Premier for routing — a wide cost spread between Micro and Premier means mis-tiered defaults burn budget silently.
• Nova Reel for long narratives — plan multi-clip generation and stitching; single-prompt long video is not the model’s design center.
• Ignoring regional model availability — Nova model IDs differ by region; hardcode in config with region checks.

Official references

• What is Amazon Nova?
• Nova 2 Lite model card
• Supported models in Amazon Bedrock
• Inference parameters for Nova models
• Amazon Nova Canvas

Related FactualMinds content

• Amazon Bedrock Consulting
• Generative AI on AWS
• Amazon Bedrock