Nova
Quick Facts
- Vendor
- Amazon (AWS)
- Released
- December 2024 (re:Invent)
- Current line
- Nova Pro · Nova Lite · Nova Micro · Nova Canvas (image) · Nova Reel (video)
- License
- Proprietary; Bedrock-hosted only
- Hosting
- Amazon Bedrock (all AWS regions supporting it)
- Context window
- 300K tokens (Pro / Lite); 128K (Micro)
- Modalities
- Text, image, video input/output depending on variant
- Positioning
- Cost-per-token competitive; frontier-adjacent quality
Summary
Nova is Amazon's first real attempt at owning the frontier-tier LLM spot in Bedrock. Before Nova, AWS customers defaulted to Claude (Anthropic) or Llama (Meta) for high-end tenancy; Amazon Titan existed but wasn't competitive. Nova changes that: Pro is benchmark-competitive with mid-tier frontier models at a fraction of the cost, Lite and Micro aggressively price the commodity tiers, and Canvas / Reel cover image and video generation so customers don't have to stitch in third-party providers.
For infrastructure teams already committed to AWS, Nova removes friction — the same IAM, VPC endpoints, CloudWatch, and KMS that govern the rest of Bedrock govern Nova too. Data residency follows AWS region settings. The tradeoff is lock-in: none of this transfers off Bedrock.
Model Lineup
- Nova Pro — flagship text. Multimodal input, 300K context, reasoning-capable.
- Nova Lite — mid-tier. Same multimodal capability at a fraction of Pro pricing; the workhorse for cost-sensitive agent loops.
- Nova Micro — text-only, optimized for latency and throughput. Routing and classification.
- Nova Canvas — image generation. Bedrock-integrated alternative to Stability and Titan Image.
- Nova Reel — video generation. Short-form clips for marketing and media.
Where Nova Fits
Nova is the default when AWS is the architectural center of gravity and cost per token is a live constraint. For agent workloads that run at high volume inside an AWS environment, Nova Lite often beats Claude and GPT on cost while staying close enough in quality. Multi-model orgs tend to route by workload — Claude for hard reasoning, Nova Lite for the bulk of cheap calls, Micro for classification.
Tradeoffs
- AWS-only. No self-hosting. If you ever move off Bedrock, you move off Nova too.
- Benchmark ceiling. Pro is frontier-adjacent, not frontier. For the hardest reasoning and coding workloads, Claude Opus and GPT-5 still win head-to-head.
- Ecosystem maturity. Third-party tooling (agent frameworks, evaluation harnesses, quantization) was built for Claude / GPT first. Most things work via Bedrock, but not all.
- Less public research. Unlike Anthropic / OpenAI / DeepMind, Amazon publishes relatively little about Nova's architecture and training, which complicates deep evaluation.
Deployment Notes
Within the Claw ecosystem, Nova is a first-class provider in the Bedrock arbitrage lane — Nova Lite often sits alongside Claude Haiku and Gemini Flash as a cheap, fast tier that soaks up routing, classification, and high-volume summarization. For AWS-native customers where data sovereignty is tied to AWS regions, Nova is usually the lowest-friction option. For on-premise Claw deployments (OpenClaw, NanoClaw), Nova doesn't apply — it lives exclusively on the cloud side of the architecture.