Amazon Nova — the complete 2026 guide.

IWhat Amazon Nova is — and AWS's positioning

Amazon Nova is Amazon's own family of foundation models, delivered as managed models inside Amazon Bedrock. Where most of the headline models on Bedrock come from partner labs — Anthropic's Claude, Meta's Llama, Mistral, Cohere — Nova is the model family AWS itself builds and trains, and it is the centre of gravity for the cost-conscious end of the Bedrock catalogue.

The simplest way to place Nova: it is to Bedrock what a house brand is to a well-stocked store. The store still sells every premium label (you can call Claude or Llama through the same API), but the house brand is engineered to deliver most of the value at a meaningfully lower price — and, importantly, with the tight integration and predictability that come from AWS owning the whole stack. Nova models run on AWS's own silicon and infrastructure, which is part of how they hit their price and latency targets.

AWS's stated positioning for Nova is explicit and worth taking at face value: price-performance and low latency, not "we beat the frontier on every benchmark." Amazon's claim is that for a large share of real production tasks, a Nova tier delivers the quality you actually need at a fraction of the cost and latency of a top frontier model — and that when you genuinely need the frontier, you call Claude (or Nova Premier) through the very same Bedrock API. This is a deliberately different pitch from "the smartest model in the world," and it shapes where Nova fits.

Nova spans multiple modalities. The core text/multimodal tiers (Micro, Lite, Pro, Premier) handle understanding tasks — reading text, images, documents and video and producing text. Separate models cover generation: Nova Canvas creates images from text, Nova Reel creates short video, and Nova Act is an agentic model designed to take actions in a web browser. So "Amazon Nova" is not one model but a small family, each member tuned for a different job.

Crucially, Nova inherits the Bedrock guarantees that matter to companies: your prompts and data are not used to train the base models, data stays in your AWS account and region, and the usual enterprise controls (IAM, PrivateLink, KMS, CloudWatch, Guardrails) apply unchanged. Choosing Nova does not mean leaving the security and governance model of Bedrock — it is the same service, the same API, a different model ID.

One caveat, stated once and meant throughout: the model names, capabilities and especially the prices on this page are representative as of 2026. AWS iterates the Nova family quickly and foundation-model prices change as the market moves. Treat the figures here as a guide to the shape of the family and its relative cost — and confirm current model availability and pricing on the official AWS Bedrock and Nova pages (and see the amazon-nova-pricing sibling for the cost detail).

IIThe Nova tiers — Micro, Lite, Pro, Premier (and Canvas, Reel, Act)

The understanding tiers form a clean ladder from cheapest-and-fastest to most-capable. The trick to using Nova well is choosing the lowest tier that clears the quality bar for each specific task — and knowing that the creative and agentic models are separate, purpose-built members of the same family.

Read the text tiers as a price/capability ladder. As you move up — Micro → Lite → Pro → Premier — the model gets more capable, handles more modalities, and costs more per token. Move down and it gets cheaper and faster. Most production systems end up using two or three tiers at once: a small tier for the high-volume easy work and a larger one for the hard cases.

IIINova at a glance — capability and context window by tier

A single scannable view of the family: what each model takes as input, the rough size of its context window, where it sits on cost, and the job it is built for. Use it to pick a starting tier; the per-token numbers live on the pricing page.

Context windows (how much text/content the model can consider at once) are large across the text tiers — comfortably into the hundreds of thousands of tokens — which makes Nova practical for long documents and big RAG contexts. The exact maximum and the supported input types evolve; the figures below are representative as of 2026.

IVHow good is Nova? An honest read vs Claude and GPT

The fair, hedged answer: Nova is genuinely strong for its price class and excellent on the broad middle of real tasks, but it is positioned as price-performance rather than frontier — and on the hardest reasoning, nuanced writing and complex code, dedicated frontier models such as Claude and GPT still tend to lead. Benchmarks are a rough guide, not gospel.

Start with the caveat that matters most: public benchmark numbers move constantly and rarely match your workload. Leaderboards (MMLU-style knowledge, math, coding suites, multimodal tests) are useful for a coarse ranking but routinely mislead on the specific task you care about. The only benchmark that truly counts is your own evaluation set on your own prompts. Treat everything below as direction, not a verdict, and run a quick eval (Bedrock has a model-evaluation feature) before committing.

With that said, here is the honest shape as of 2026. On the broad middle of production work — classification, extraction, retrieval-augmented answers, summarization, structured/JSON output, routine tool calling, straightforward multimodal understanding — the right Nova tier is typically good enough to ship, and it gets there far cheaper and faster than a frontier model. For a great many real features, you would struggle to tell a well-prompted Nova Pro answer from a frontier-model answer, and you would pay a fraction of the cost.

On the hard end — multi-step reasoning, subtle instruction-following, long-horizon agentic planning, nuanced or stylistically demanding writing, and especially difficult code generation and debugging — dedicated frontier models still tend to have an edge. Anthropic's Claude in particular is widely regarded as a leader for complex reasoning and coding, and is available on the very same Bedrock API; OpenAI's GPT models (accessed outside Bedrock) are the other common frontier reference point. Nova Premier narrows this gap and is the family's answer for hard tasks, but "narrows" is the honest word, not "erases."

The productive way to hold this is not "Nova vs Claude" as a one-time pick but as a portfolio. A tiered router sends the easy 70–90% of requests to a cheap Nova tier and escalates only the genuinely hard minority to Nova Premier or Claude. Because both live behind one Bedrock API, this routing is a software decision, not an integration project — and it tends to cut cost dramatically while keeping quality high where it matters. That is exactly the architecture Nova is designed to enable.

VHow to access Nova — Bedrock, model IDs, and the API

Every Nova model is reached through Amazon Bedrock, so if you can call any model on Bedrock you can call Nova by changing one identifier. There is no separate Nova endpoint, SDK or account to set up — it is the same managed service with a different model ID.

The workflow is the standard Bedrock workflow. First, in the Bedrock console under Model access, enable the Nova models you want in your region (a one-time toggle). Then call them with the AWS SDK (boto3, the AWS SDK for JavaScript, etc.) using either the Converse API — the unified, recommended interface for chat-style and multimodal requests, with a consistent message format across all models — or the lower-level InvokeModel call. Switching from, say, Claude to Nova Pro is changing the modelId string; the surrounding code stays the same.

The model IDs follow the pattern amazon.nova-<tier>-v1:0: amazon.nova-micro-v1:0, amazon.nova-lite-v1:0, amazon.nova-pro-v1:0, amazon.nova-premier-v1:0, with corresponding IDs for nova-canvas and nova-reel. For higher availability and to avoid single-region throttling, many teams call Nova through a cross-region inference profile (an "inference profile ID" that lets Bedrock serve the request from a pool of regions) — see the amazon-bedrock-cross-region-inference sibling.

Because Nova lives inside Bedrock, every surrounding Bedrock capability works with it unchanged: Knowledge Bases for managed RAG, Agents for tool-using workflows, Guardrails for safety/PII filtering, Prompt Management and Flows for orchestration, and fine-tuning / distillation to customise a model. You can also pick a pricing mode per workload — on-demand, Batch (~50% cheaper for async jobs), Provisioned Throughput, and prompt caching — exactly as you would for any Bedrock model (the amazon-bedrock-pricing sibling covers the modes).

• Enable access — Bedrock console → Model access → enable the Nova models in your region (one-time).
• Call via Converse API — Use the unified Converse API with modelId = amazon.nova-pro-v1:0 (or the tier you want); the same code calls any model by swapping the ID.
• Add cross-region inference — Use an inference-profile ID to spread load across regions and reduce throttling for production traffic.
• Layer Bedrock features — Attach Knowledge Bases (RAG), Agents, Guardrails, prompt caching and Batch exactly as with any Bedrock model.
• Govern as usual — IAM, PrivateLink, KMS, CloudWatch and the no-training-on-your-data guarantee all apply unchanged.

VIUse cases by tier — what to run where

Nova rewards the discipline of matching the tier to the task. Spend a frontier-class budget only where the task is genuinely hard; push everything else down the ladder. Here is a practical mapping of common workloads to the tier that usually fits.

These are starting points, not laws — always confirm with a quick eval. The recurring theme is that a great deal of production GenAI work is "easy" in model terms (well-defined, repetitive, schema-bound), and that work belongs on Micro or Lite, with Pro as the balanced default and Premier (or Claude) reserved for the hard minority.

• Nova Micro — high-volume simple text — Intent classification and routing, sentiment/topic tagging, entity and field extraction, short summaries, content moderation pre-filtering, function/tool calling, the cheap first pass in a router, and any text task done at massive volume where latency and cost dominate.
• Nova Lite — cheap multimodal at scale — Image classification and description, reading documents/screenshots/receipts/forms, multimodal RAG, light video understanding, product-catalog enrichment, and high-volume "understand this image/doc" features that need vision without frontier prices.
• Nova Pro — the balanced default — The main model for many apps: agent orchestration and tool use, richer document and video analysis, complex structured/JSON output, customer-facing assistants where quality matters but cost still counts, and anything where Lite is not quite reliable enough.
• Nova Premier — hardest tasks + distillation — Demanding multi-step reasoning, complex agentic planning, high-stakes analysis, and as the teacher model to distill a smaller, cheaper custom model for a narrow high-volume task. (For the very hardest reasoning/code, also evaluate Claude.)
• Nova Canvas — images — Marketing and product imagery, design exploration, image editing/inpainting, and in-app image generation where you want it to run inside AWS with provenance safeguards.
• Nova Reel — short video — Social and marketing clips, b-roll, motion concept prototyping, and image-to-video without a production shoot.
• Nova Act — browser automation — Agents that operate real web UIs: research, form filling, multi-step workflows across sites without good APIs. Treat as an emerging capability to prototype against.

VIIWhen Nova vs Claude vs Titan

Three "Amazon-adjacent" choices confuse people: Amazon Nova (the current price-performance family), Anthropic's Claude (the frontier reasoning/coding leader, also on Bedrock), and Amazon Titan (the earlier Amazon model family). Here is how to choose between them in 2026.

Nova vs Claude. This is the live, everyday decision, and the honest framing is "value vs frontier." Reach for Nova when cost and latency matter and the task sits in the broad middle — classification, extraction, RAG answers, summarization, structured output, high-volume agents, multimodal understanding. Reach for Claude when the task is genuinely hard — complex reasoning, nuanced or long-form writing, difficult code, subtle instruction-following — where its quality edge justifies the higher price. Because both are on Bedrock behind one API, the best answer is usually both via a router: Nova for the easy majority, Claude for the hard minority. See the claude-on-amazon-bedrock sibling for the Claude side.

Nova vs Titan. Amazon Titan was Amazon's first-generation foundation-model family on Bedrock. For new generative text and multimodal work in 2026, Nova is the successor and the better choice on quality, capability and price-performance — if you are starting fresh, default to Nova, not Titan, for generation. Titan's most enduring relevance is its embeddings models: Amazon Titan Text Embeddings remains a common, cheap, solid choice for the retrieval/vector half of a RAG or search system. So a very typical modern stack is "Nova for generation, Titan Text Embeddings for retrieval." See the amazon-titan sibling for detail.

A simple rule of thumb. Default to Nova for new generation work; escalate to Claude for the hard cases (ideally via a router on the same API); use Titan Text Embeddings for embeddings/RAG retrieval. You are not locked in to any of these — switching among them is a model-ID change on Bedrock, so the right move is to start with Nova, measure on your own eval set, and escalate only where the data says you must.

VIIIGetting started with Nova — and making it $0

Getting from "interested in Nova" to a running, cost-tuned workload is short, because Nova is just Bedrock with a different model ID. The only meaningful decision beyond which tier is whether to pay for it yourself or have AWS credits cover the build — which, for most startups and many companies, they will.

The mechanical first steps: (1) enable the Nova models you want under Bedrock Model access; (2) prototype against Nova Pro as a sensible default with the Converse API; (3) build a small evaluation set from your real prompts and compare Pro against Lite/Micro (to push cost down) and against Premier or Claude (where you need more); (4) introduce a tiered router once you know which requests are easy vs hard; and (5) turn on the cost levers that fit — Batch for bulk jobs, prompt caching for repeated context, Provisioned Throughput only for steady high volume. That sequence gets most teams to a cheap, reliable production setup quickly.

The cost story is where CloudRoute comes in. Nova is the value choice on Bedrock, but at real scale GenAI still costs money — and AWS will frequently fund the build with credits. Nova inference, fine-tuning, distillation and the supporting services are all credit-eligible. The relevant pools: AWS Activate (general startup credits, commonly up to $100K for institutionally-funded startups), a dedicated Bedrock / Generative-AI POC pool ($10K–$50K) for proving out a use case, and the competitive Generative AI Accelerator (awards up to $1M for a small cohort of AI-first startups). Credits apply automatically against your AWS bill — including Nova usage — until exhausted.

Most of those pools are partner-filed through the AWS Partner Network (the ACE program), not a public self-serve form, which is why teams route through an AWS partner rather than applying alone. That is the gap CloudRoute fills: it matches you to the right credit pool for your stage and to a vetted AWS DevOps/ML partner who both files the credit application and helps build the Nova workload (the RAG pipeline, the agent, the tiered router, the distillation). The customer pays $0 — AWS funds the credit pool, AWS pays the partner through engagement-funding programs, and the partner pays CloudRoute a routing commission. You never see an invoice. For the credit mechanics specifically, see the cross-cluster pages on AWS credits for generative-AI startups and Bedrock POC funding.