WatchModels & PlatformsNo changeMarch 2026 Backfill

Early or unproven. Follow the space but don't commit yet.

Amazon Nova

Competitive on price-performance for Bedrock-centric enterprises, but trails frontier models on reasoning and coding and remains locked to the AWS ecosystem.

LLM·Multimodal

aws.amazon.com

Our Take

What It Is

Amazon Nova is AWS's homegrown foundation model family. The Nova 2 generation, launched at re:Invent December 2025, includes four models: Lite (cost-effective reasoning, 1M context), Pro (complex multi-step tasks, still in Preview), Omni (multimodal reasoning across text, images, video, and speech), and Sonic (real-time speech-to-speech). All models run exclusively on Amazon Bedrock.

Why It Matters

Nova sits at Watch because the models are competitive on price-performance within the AWS ecosystem but aren't differentiated enough to pull teams away from Anthropic, OpenAI, or Google. Nova 2 Lite at $0.30/M input tokens is genuinely cheap, and the Bedrock integration means zero additional infrastructure for AWS-native organisations.

The practical consideration: if your organisation already runs on AWS and needs cost-effective AI capabilities with integrated governance, Nova is worth evaluating. If you need frontier reasoning or coding performance, you're looking elsewhere.

Key Developments

  • Dec 2025: Nova 2 family launched at re:Invent — Lite, Pro (Preview), Omni, Sonic.
  • Dec 2025: Nova Forge announced for customers to build custom frontier models on AWS infrastructure.
  • Dec 2025: Nova Act announced for browser-based UI workflow automation (computer use).
  • Dec 2025: Nova Multimodal Embeddings — first unified model for text, documents, images, video, and audio.

What to Watch

Pro leaving Preview and reaching GA is the key milestone. If Nova 2 Pro matches Gemini 2.5 Pro on reasoning and coding when it reaches production, the combination of AWS integration and competitive pricing makes a strong case for Bedrock-native teams. Watch for third-party benchmarks — AWS's own comparisons should be treated with healthy scepticism.

Strengths

  • Aggressive price-performance: Nova 2 Lite at $0.30/M input tokens, positioned as the budget reasoning option in the market.
  • Broad multimodal coverage: Nova 2 Omni processes text, images, video, and speech while generating both text and images natively.
  • Deep AWS integration: Seamless through Bedrock with cross-region inference, IAM, and governance tooling.
  • 1M-token context window: Nova 2 Lite with developer controls for thinking effort and built-in tools.

Considerations

  • AWS ecosystem lock-in: All Nova models available exclusively through Bedrock. No self-hosting or non-AWS deployment.
  • Reasoning and coding trail frontier: Weaker on advanced math, science, and coding compared to Gemini 2.5 Pro and o3.
  • Pro still in Preview: Not yet GA for production workloads as of March 2026.
  • Latency: Nova Pro processes ~95 tokens/second, slower than GPT-4o at ~115 tokens/second.