What it is
Nova is Amazon's own family of AI models. Amazon is best known as a shop, but it also runs Amazon Web Services (AWS), one of the world's largest cloud businesses, which rents computing power and software to companies over the internet. Nova models live there. A business pays AWS to send text, pictures, video or speech to a Nova model and get an answer back, and Amazon uses the same models inside its own products.
Like GPT, Claude and Gemini, the text models in the family are language models: systems trained on huge amounts of writing so that they can read, summarise, answer questions and write. Amazon also built Nova models that make pictures, make short videos, hold spoken conversations, and operate a web browser.
Before Nova: Titan
Nova was not Amazon's first attempt. On 13 April 2023, AWS announced Bedrock, a service that lets businesses use several companies' models through one account, and alongside it a family of in-house models called Titan. At first Titan was available only in a limited preview, and Amazon declined to say how big the models were or what data trained them. In November 2023 AWS added Titan Image Generator, which puts an invisible watermark in every picture it makes. Amazon's most visible AI bet in this period, though, was a partnership: by November 2024 it had invested a total of $8 billion in Anthropic, the maker of Claude, and made AWS Anthropic's main cloud and training partner.
The launch and the first family
Amazon announced Nova on 3 December 2024 at re:Invent, its yearly cloud conference in Las Vegas. Amazon's chief executive, Andy Jassy, made a surprise appearance on stage to introduce it. The work came from Amazon's artificial general intelligence group, then led by Rohit Prasad. Five models were available that day:
- Nova Micro, a fast, low-cost model that handles text only.
- Nova Lite and Nova Pro, which can read text, images and video and reply in text. Pro was the stronger and more expensive of the two.
- Nova Canvas, which creates images from a description. Like Titan's image model, it adds an invisible watermark, and it also attaches Content Credentials, a labelling standard that records which tool made a picture.
- Nova Reel, which creates short video clips. At launch the clips were six seconds long.
Amazon's own claim at launch was that the models were at least 75 percent cheaper than the best-performing models in their class on Bedrock. That is the company's comparison, not an independent test. A larger model, Nova Premier, followed on 30 April 2025, and a voice model, Nova Sonic, on 8 April 2025. Sonic is a speech-to-speech model: you talk, it listens and talks back, without first turning your voice into written text and back again.
The second generation
On 2 December 2025, again at re:Invent, Amazon announced Nova 2. Nova 2 Lite became available that day as an everyday model that can "think" before answering, at three levels of effort you can choose (see Extended Thinking). Nova 2 Pro, the most capable, and Nova 2 Omni, a model that takes in text, pictures, video and speech and can reply with text or images, launched in preview, meaning early testing rather than full release. Nova 2 Sonic replaced the first voice model. The text models can take in up to a million tokens, the small chunks of text models read, which is roughly several long books at once (see Context Window).
The same event brought two related products. Nova Forge lets a large company train its own version of Nova, mixing its private data into copies of the model saved at different stages of training. SiliconANGLE reported that subscriptions started at around $100,000 a year. Nova Act, a tool for building agents that use websites, moved from a research preview to a full AWS service.
Proprietary and cloud-hosted
This is the most important thing to understand about Nova if you are not technical. Nova is proprietary: Amazon owns it and keeps it on its own computers. You cannot download it. Some companies, such as Meta with Llama and Mistral with some of its models, release "open weights", the learned numbers inside a model (its parameters), so that anyone can run a copy on their own machine. Nova is not like that.
In practice this means three things. First, you always reach Nova through Amazon, either on its website or through the AWS API, a way for one program to use another, and you pay per use. Second, whatever you type travels to Amazon's servers, so what happens to it depends on Amazon's terms. Third, Amazon decides when a model changes or disappears. If it retires a model, everyone who built on it has to move, which, as the examples below show, is not a theoretical risk.
Why it matters
It shows how the cloud giants think. Microsoft and Google rent out other companies' models and also build their own. Amazon does the same. Bedrock sells models from Anthropic, Meta, Mistral and others (the Mistral and Qwen pages mention theirs), and Nova sits beside them as Amazon's in-house choice. The pitch is price and convenience for businesses already on AWS, rather than being the single smartest model in the world.
You may be using it without knowing. Most people will never type "Nova" into anything. They meet it through Alexa, through customer service phone lines built on AWS, or through apps whose makers picked a Nova model because it was cheap enough to run millions of times a day. Knowing that a household name's assistant may switch between models from different companies helps explain why answers can vary in quality and style.
It is a live lesson in how fast this field moves. In less than two years Nova went from launch to a second generation to Amazon retiring several of its models. For anyone choosing an AI tool for a business, that is a reminder to plan for change: keep your prompts and data portable, and don't assume the model you picked this year will exist next year.
Real-world examples
Alexa+ and the question of which model is talking
On 26 February 2025, Panos Panay, Amazon's head of devices and services, unveiled Alexa+ at an event in New York: a rebuilt, generative-AI version of the Alexa voice assistant, priced at $19.99 a month or free with Amazon Prime. Amazon said it drew on models from several providers, including its own Nova and Anthropic's Claude, and Anthropic announced the same day that Claude models were "helping power Alexa+".
Two days later, on 28 February, CNBC reported, citing unnamed sources, that Claude was handling most of the questions customers were asking the new Alexa. Amazon called the report false. A spokesperson told CNBC that "over the past four weeks Nova handled over 70% of conversations", including complex requests. Those are two different claims from two sides, and outsiders had no way to check either. What is clear is Amazon's own description: Alexa+ picks whichever model it judges best for a request. On 4 February 2026, Amazon made Alexa+ available to everyone in the US, free for Prime members, $19.99 a month for others, and with a limited free text chat at Alexa.com. Its announcement describes the assistant as powered by "large language models—from both Amazon Nova and Anthropic". Prices and plans may have changed since, so check Amazon's current page.
Nova Act: an agent that uses the web
On 31 March 2025, Amazon released a research preview of Nova Act, a model trained to operate a web browser, together with a software kit for developers and a new website, nova.amazon.com, where people in the US with an Amazon account could try Nova models. TechCrunch reported that it came from Amazon's San Francisco AGI lab, co-led by David Luan and Pieter Abbeel, and that Amazon said it would power features in Alexa+.
Amazon's approach was to break jobs into small, reliable steps, such as "search for this", "choose that date", or "fill in this form", instead of handing an agent one vague instruction. Amazon reported a 94 percent score on ScreenSpot Web Text, a test of whether a model can find the right thing to click, ahead of the scores it reported for OpenAI's and Anthropic's agents. TechCrunch noted that Amazon did not report results on WebVoyager, a more common agent test, and that agents of this kind in general were slow and error-prone. When Nova Act became a full AWS service in December 2025, Amazon said it delivered "90% reliability on early customer workflows" and named Hertz among its users. Those are the company's figures from its own customers. See Browser Agents for what these tools can and can't do.
The 2026 wind-down
On 28 July 2026, Business Insider reported, and Reuters followed, that Amazon was winding down most of its flagship Nova models to concentrate on a new frontier model. By then Amazon had already reorganised: Rohit Prasad left at the end of 2025, Peter DeSantis took over a combined AI, chips and quantum computing group, and Pieter Abbeel was put in charge of frontier model research. According to the reports, Premier, Omni, Reel and Canvas were moving to maintenance only, while Nova 2 Lite, Nova 2 Sonic, Nova Forge and Nova Act would continue. An Amazon spokesperson said AI models "remain one of the most important things we're working on". A new flagship model was expected at re:Invent later in 2026.
Amazon's own documentation shows what that means for customers. Bedrock's model lifecycle table lists Nova Premier and the first Nova Sonic with an end-of-life date of 14 September 2026, and Nova Canvas and Nova Reel with 30 September 2026. After that date, requests to those models fail and customers must move to something else themselves. For a business that built a product on Nova Canvas in 2025, that is roughly eighteen months from launch to shutdown.
An everyday case: a shop owner picks a model
This is an illustration, not a documented case. Priya runs a small online homeware shop whose website already runs on AWS. Her developer suggests a chatbot that answers questions like "Will this rug fit a 2-metre hallway?" using the shop's product sheets. They test a Nova model and two others on Bedrock with the same fifty real customer questions. Nova is the cheapest, and good enough on straightforward questions, but it guesses on two questions about return rules, so they give it the returns policy word for word and tell it to say "I'm not sure, please email us" when the answer isn't there. They also write the prompts so that they could switch to another model on Bedrock in an afternoon, because they have read that models get retired.
Try it yourself
Websites and menus change often, and at the time of writing some features are limited to the US. Treat these steps as a guide and check what's offered today.
- Go to nova.amazon.com. At the time of writing it lets you try current Nova models, including Nova 2 Lite for chat and Nova Act. You may be asked to sign in with an ordinary Amazon account. If the site isn't available where you live, the free text chat at Alexa.com (also Amazon account, US at the time of writing) is another way to meet Amazon's assistant, though it may use Nova or Anthropic models without saying which.
- Give it a real everyday task. Paste something you would actually use, for example: > Here is an email from my landlord. Summarise what he is asking me to do, by when, and list any questions I should ask him before I reply. [paste the email]
- Test it on something you can check. Ask about a topic you know well, such as your job, your town, or a hobby. Note anything that is wrong or made up. All language models do this sometimes; see Hallucination.
- Compare with another model. Paste the same prompt into a chatbot from another company, such as ChatGPT, Claude or Gemini. Don't ask which is "best". Ask which answer was more accurate, clearer, and more useful for this particular job.
- Try the thinking setting, if offered. Ask a small logic puzzle or a planning question ("Plan a three-day trip to Edinburgh by train on a £300 budget") with thinking off and then on. Notice the difference in time and in quality.
- Read before you paste anything private. Look for the site's privacy notice and terms. Remember that Nova runs on Amazon's computers, not yours, so don't paste passwords, medical details, or other people's personal information. See Privacy.
- Check which models still exist. Open Amazon's Bedrock model lifecycle page (in the resources below) and see which Nova models are listed as legacy or end-of-life. It is a useful habit before relying on any model name you read about, including in this article.
Common mistakes and misunderstandings
"Nova is a chatbot app like ChatGPT." Mainly it isn't. Nova is a set of models sold to businesses through AWS, with a website for trying them. The assistant most consumers meet is Alexa+, which uses Nova alongside models from Anthropic.
"Alexa+ is just Nova" or "Alexa+ is just Claude". Amazon says Alexa+ uses both and picks a model for each request. Reports have disagreed about how the work is split, and the split can change at any time without users being told.
"Nova is open source because Amazon offers Forge." Nova Forge gives paying companies access to saved stages of Nova's training so they can build their own version on AWS. It doesn't make the weights public. You cannot download Nova and run it on your own computer the way you can with open-weight models such as Llama or Gemma.
Treating launch numbers as independent results. "75 percent cheaper", "equal or better on 13 out of 15 benchmarks", and "90% reliability" are Amazon's own claims, measured by Amazon, on tests it chose. Other companies do the same with their models. Treat any benchmark quoted by a model's maker as a starting point, and test on your own work.
Assuming a model you read about still exists. Nova Premier, Canvas and Reel were being retired in September 2026, less than two years after the family launched. Product names, versions, prices and availability change quickly, so check Amazon's current model list before relying on any of them.
Confusing Nova with Titan. Titan was Amazon's earlier model family from 2023, which included text models, an image generator, and embedding models (which turn text and pictures into numbers for search, see Embeddings). Nova is the newer line. If a guide or tutorial talks about Titan, it may be out of date.
Thinking the picture and video tools are the main event. Canvas and Reel got headlines at launch, but Amazon's continued effort, according to the 2026 reports, is on the text, voice and agent models. If you want AI images or video, dedicated tools covered on the Diffusion Model page and in the creative tools row are the more common choice.
Related elements
- Claude: Anthropic's models, which Amazon has invested in, sells on Bedrock, and uses alongside Nova in Alexa+.
- GPT and Gemini: the model families from OpenAI and Google that Nova is measured against.
- Llama: Meta's open-weight models, a useful contrast to Nova's cloud-only approach.
- Parameters: the learned numbers that open-weight models share and proprietary models like Nova keep private.
- Browser Agents: the kind of tool Nova Act is built for.
- Computer Use: AI that operates screens and apps, a close cousin of Nova Act.
- Extended Thinking: the adjustable "thinking" built into Nova 2 Lite and Pro.
- Fine-tuning: adapting a model with your own examples, which Nova Forge takes further.
- Customer Support: where voice models like Nova Sonic answer phone calls.
- Benchmarks: why a company's own scores need a second look.
- API: the way businesses connect their software to Nova through AWS.