Amazon Nova: Amazon's own AI models on Bedrock.

IWhat Amazon Nova is

Amazon Nova is the model family AWS builds and trains itself, served as managed models inside Amazon Bedrock. Most other models on Bedrock come from partner labs: Anthropic, Meta, Mistral, Cohere and others.

AWS positions Nova on price-performance and low latency. The claim is that for a large share of production tasks a Nova model gives the quality you need at a fraction of a frontier model's price, and that when you need the frontier you can call another model through the same Bedrock API.

Nova is not one model. AWS describes it as a portfolio: Nova models (text, multimodal, speech and embeddings), Nova Act (a service for building agents that automate browser-based workflows) and Nova Forge (a service for building your own model on top of Nova).

Nova runs inside Bedrock, so the usual controls apply: IAM, PrivateLink, KMS, CloudWatch and Guardrails. AWS says Bedrock does not use your prompts or outputs to train the base models.

Model names and prices change often. The prices on this page are AWS list prices for us-east-1 on-demand, checked on October 10, 2026. Confirm on the AWS Bedrock pricing page for your Region.

IIThe Nova lineup in October 2026

Pick the cheapest model that clears your quality bar for each task. Most production systems use two or three models at once.

IIINova models at a glance: inputs, context and price

One view of the text and multimodal models: what each takes in, how much context it holds and what it costs per 1 million tokens (AWS list price, us-east-1 on-demand, checked October 10, 2026).

IVHow good is Nova compared with Claude and GPT?

Nova is strong for its price and good enough for a lot of real work. On the hardest reasoning, careful writing and complex code, frontier models such as Claude and GPT still tend to lead.

Public benchmarks rarely match your workload, so build a small test set from your own prompts and use Bedrock's model evaluation feature before you commit.

On the broad middle of production work (classification, extraction, RAG answers, summaries, structured output, routine tool calls, image and document reading) the right Nova model is usually good enough, and cheaper and faster than a frontier model.

On the hard end, Claude Sonnet and Opus on Bedrock are the models to test, and OpenAI's GPT models are the other frontier reference outside Bedrock. Price is no longer one-sided: at October 2026 list prices, Claude Haiku 5.5 costs less per token than Nova Pro and Nova 2 Lite. See Amazon Nova vs Claude on Bedrock for the full comparison.

The practical pattern is a router: a cheap model for the easy majority, a frontier model for the hard minority, all behind one Bedrock API.

VHow to access Nova on Bedrock

Every Nova model is reached through Amazon Bedrock. If you can call any Bedrock model, you can call Nova by changing one identifier.

• Check what is active: In the Bedrock console, check which Nova models are available in your Region; availability differs by Region and model version.
• Call it with the Converse API: Use the unified Converse API (or InvokeModel) from the AWS SDK. First-generation IDs look like amazon.nova-pro-v1:0; the same code calls any model by swapping the ID.
• Use cross-region inference: An inference profile ID (for example us.amazon.nova-pro-v1:0) spreads requests across Regions and reduces throttling.
• Add Bedrock features: Knowledge Bases for RAG, Agents, Guardrails, prompt caching and Batch work with Nova as with any Bedrock model.
• Customize if you need to: Fine-tune or distill a Nova model on Bedrock or SageMaker AI, or use Nova Forge to build your own model from Nova.

VIUse cases: which Nova model where

Starting points, not rules. Confirm with a quick test on your own prompts.

• Nova Micro: high-volume simple text: Intent classification, routing, tagging, field extraction, short summaries and the cheap first pass in a router.
• Nova Lite: cheap multimodal at scale: Describing or classifying images, reading documents, screenshots and forms, and light video understanding.
• Nova 2 Lite: everyday reasoning and long context: Agents and tool use, long documents up to 1 million tokens, and tasks that benefit from thinking turned up.
• Nova Pro and Premier: first-gen balanced and top tiers: Existing workloads built on them, and Premier as a distillation teacher for a narrow, high-volume task.
• Nova 2 Sonic: voice: Voice agents, customer service lines and real-time spoken conversations.
• Nova Multimodal Embeddings: search: Semantic search and RAG across text, images, video and audio.
• Nova Act: browser agents: Agents that complete multi-step tasks in web interfaces that lack good APIs.

VIIWhen to use Nova, Claude or Titan

Three Bedrock options confuse people: Amazon Nova, Anthropic's Claude and Amazon Titan, Amazon's earlier model family.

Nova vs Claude. Use the cheapest model that passes your tests for cost-sensitive, high-volume and latency-sensitive work (Nova Micro, Nova Lite, or Claude Haiku 5.5). Use Claude Sonnet or Opus for hard reasoning, careful writing and real code. Both are on Bedrock behind one API, so a router can use both.

Nova vs Titan. For text generation, use Nova, not Titan. AWS's current Bedrock model-card list includes the Titan embeddings models but not Titan Text or Titan Image Generator, and the old Titan product page now opens the Bedrock page. Titan Text Embeddings V2 is still a common, low-cost choice for retrieval.

Rule of thumb. Nova for new generation work, Claude for the hard cases, and Titan Text Embeddings V2 or Nova Multimodal Embeddings for retrieval. Switching is a model-ID change, so start cheap, measure and escalate only where the data says so.

VIIIGetting started, and who files which AWS funding

Getting from "interested in Nova" to a running workload is short, because Nova is Bedrock with a different model ID. The other question is who pays for the build.

First steps: (1) check which Nova models are active in your Region; (2) prototype on Nova 2 Lite with the Converse API; (3) build a small test set from real prompts and compare it with Nova Micro or Lite (to cut cost) and with Claude Sonnet (where you need more); (4) add a router once you know which requests are easy and which are hard; (5) turn on Batch for bulk jobs and prompt caching for repeated context.

Partner-filed funding (the bigger route for most startups). Nova usage is billed through AWS, so AWS money applies to it. A vetted AWS consulting partner can file proof of concept (POC) funding to test a Nova workload, or incremental workload (IW) funding if you already run on AWS and are launching new work, such as a new AI feature. For startups this is often $50K to $150K. Depends on the workload; AWS approves. The partner files it, not you.

AWS Activate credits (you apply yourself). Activate Founders is $1,000 to start, up to $5,000; you apply on the AWS site. Activate Portfolio is up to $200,000 with the Org ID from a VC, accelerator or angel group that is an Activate Provider. A consulting partner cannot file Activate for you.

CloudRoute matches you with one vetted AWS partner who can build the Nova workload (the RAG pipeline, the agent, the router) and file the partner funding when the work qualifies. Startups pay CloudRoute nothing. Partners pay CloudRoute a monthly membership plus 10% of AWS-approved funding.