Amazon Nova

What it actually is

Nova is Amazon’s answer to having its own AI models instead of only reselling other companies’ models on Bedrock. It launched at the end of 2024 with four core models — Micro for simple text tasks, Lite for fast multimodal work, Pro for a stronger all-around option, and Premier for harder reasoning — plus Canvas for images and Reel for video. A year later, Amazon expanded the family further with Nova 2, adding a real-time voice model and Nova Act, a system built specifically to look at a screen and operate it the way a person would, clicking and filling in forms rather than just answering questions about text.

None of this is sold as a product you open and use directly. It’s sold as infrastructure — every Nova model lives inside Amazon Bedrock, priced per token, the same way you’d pay for any other model hosted there. Amazon uses the same models inside its own products, including Rufus, the shopping assistant built into the Amazon app, and parts of Alexa+.

Where it’s genuinely good

Cost is the real story, and it’s nowhere near. Nova Micro’s price per token undercuts every comparable model from any other major cloud provider, and for the kind of task that shows up millions of times a day in a real system — classifying a message, pulling a name out of a form, answering a simple question — that difference adds up fast. Teams building large systems on Bedrock often route the easy, high-volume work to Nova and save a more expensive, more capable model for harder cases, keeping the whole system affordable without sacrificing quality where it matters.

Nova Act is the more surprising strength. Getting an AI system to reliably read a real, messy website or app and actually operate it — click the right button, fill in the right field — is a harder problem than it sounds, and independent testing has found Nova’s models notably better at this specific skill than comparable models built by other companies. Combined with real-time voice and video generation in the same family, Amazon has built something with an unusually wide range for one connected lineup.

Where it falls short

Getting started is the first real hurdle. There’s no app to download and no simple monthly plan — using Nova means opening an AWS account and learning Bedrock, which has its own real learning curve around regions, permissions, and several different ways of being billed. That’s a fair trade for a company already running on AWS. It’s a genuine barrier for anyone who just wants to try an AI model without becoming a cloud administrator first.

The capability ceiling is the other honest limitation. Nova’s models are good, sometimes very good for their price, but independent testing consistently places them a step behind the strongest models from the top labs once a task gets genuinely hard — deep multi-step reasoning, difficult coding problems, the kind of work where the best available model actually matters more than the cheapest one. Nova was built to win on price and breadth, not to top the leaderboard, and it’s worth choosing it with that tradeoff in mind rather than expecting frontier-level results at a fraction of the price.

Which option to actually pick

Start with Nova Micro or Nova Lite for anything simple and high-volume — classification, short answers, extracting information from text. Move up to Nova Pro once a task needs real multimodal understanding, like reading documents that mix text and images, or handling longer, more complex requests. Reach for Nova Premier when a task needs deeper reasoning but you still want to stay inside the Nova family for consistency or cost reasons — otherwise, most serious Bedrock users route their hardest tasks to a stronger model and keep Nova for everything else. If your work involves automating an actual on-screen interface rather than just generating text, Nova Act is worth testing on its own, separate from the question of which text model you’re using.

Ease of use

There’s a real gap between “easy once you’re set up” and “easy to get started.” Sending a request to a Nova model through Bedrock’s API is straightforward if you’ve built with any cloud AI service before. Getting to that point — creating an AWS account, requesting model access, understanding which region has which model available, and picking the right billing mode — is a genuine project for anyone new to AWS, not a five-minute signup like a consumer chatbot.

Who it’s actually for

Companies and developers already building on AWS who want a cheap, fast model for the high-volume, everyday parts of a larger system, and who are comfortable mixing in a stronger model from elsewhere for the harder tasks. It’s also a strong pick for anyone building something that needs to operate a real interface—clicking, scrolling, filling forms—rather than just producing text. It’s not the right choice for someone who wants a single AI model to handle everything at the highest possible quality, or anyone looking for a simple app to open and start chatting in, since neither of those is what Nova was built to be.

Pricing, model, and benchmark details were reviewed against Amazon’s official Bedrock pricing page, Amazon’s published technical model report, and independent third-party benchmarking as of August 2026. Confirm current terms before building on it—the model lineup has already undergone one full generational change since launch.