What is Amazon Nova 2, and how is it different from the first Nova release?

Nova 2 is the new generation of Amazon Nova foundation models launched at AWS re:Invent 2025. It adds four models, including Lite, Pro, Sonic and Omni, with a focus on long context, multimodal inputs and built in tools for web grounding and code execution. Compared with the first Nova release, this generation is explicitly tuned for agentic workloads and tightly integrated with Nova Forge, Nova Act and Amazon Bedrock AgentCore.

How does Nova 2 help developers build better AI agents?

Nova 2 Lite and Nova 2 Pro include reasoning controls, web grounding and code execution so agents can both plan and act, not only generate text. Nova 2 Sonic brings real time speech to speech interaction, and Nova 2 Omni can reason across text, images, video and speech while generating both text and images. Combined with Nova Forge and AgentCore, this lets developers design agents that run for long periods, coordinate tools and stay within clear safety boundaries.

What is Nova Forge, and why should enterprises care?

Nova Forge is a new service that lets enterprises build their own frontier models, called Novellas, using Amazon Nova checkpoints and proprietary data. Teams can combine pre trained, mid trained and post trained checkpoints with their own datasets, plus reinforcement learning gyms and distillation, to shape models around business specific tasks while keeping the underlying frontier capabilities. For enterprises that need control, performance and privacy, this is a way to move beyond simple fine tuning.

How do AWS Frontier agents and Nova Act fit into the overall agentic AI stack?

Frontier agents such as Kiro, AWS Security Agent and AWS DevOps Agent are packaged agents that target specific roles in the software lifecycle, from coding to security and operations. Nova Act is a managed service for browser automation agents powered by a custom Nova 2 Lite model, aimed at UI workflows like form filling and regression testing. Together with AgentCore, they form a layered stack where Nova models provide intelligence, AgentCore handles orchestration and governance, and agents deliver task and project-level outcomes.

AWS doubles down on agentic AI with Nova 2 and new agents

Updated on December 06, 2025 10 minutes read

Photorealistic data center corridor with glowing server racks and floating “AI agent” dashboards connected by neon data lines, symbolising cloud-based agentic AI.

AWS is turning its Nova foundation models into a full stack for agentic AI that combines models, customisation tools, and managed agents. At AWS re: Invent 2025 in Las Vegas on 2 December 2025, the company introduced the Nova 2 model family, the Nova Forge customisation program, the Nova Act browser agent service, and a trio of frontier agents for software teams. New policy, evaluation, and memory features in Amazon Bedrock AgentCore are intended to keep those agents within clear boundaries while they run for hours or days. For developers and learners, the focus is shifting from one-off chatbots to production-ready AI teammates that live inside the AWS ecosystem.

What happened

On 2 December 2025, AWS used its re: Invent keynote in Las Vegas to announce Amazon Nova 2, the second generation of its Nova foundation models for agentic workloads. The new family consists of Nova 2 Lite and Nova 2 Pro for text-first reasoning, Nova 2 Sonic for real-time speech, and Nova 2 Omni for unified multimodal reasoning and generation. All are accessed through Amazon Bedrock, with Nova 2 Lite and Nova 2 Sonic generally available and Nova 2 Pro and Nova 2 Omni offered through early access programs tied to Nova Forge and selected customers.

Nova 2 Lite is described as a fast, cost-effective reasoning model for everyday workloads that can take text, images, documents, and video as input and generate text outputs. It supports extended thinking modes so developers can trade off speed and depth of reasoning, offers a one-million-token context window, and ships with built-in web grounding and a code interpreter to anchor answers in current data and executable code. Benchmarks published by Amazon claim Nova 2 Lite is equal to or better on 13 of 15 tests compared with Claude Haiku 4.5, on 11 of 17 compared with GPT 5 Mini, and on 14 of 18 compared with Gemini Flash 2.5 in its size class.

Nova 2 Pro is positioned as Amazon’s most capable reasoning model for complex, multistep tasks such as agentic coding, multi-document analysis, and long-range planning. It can process text, images, video, and speech to produce text, and is explicitly designed to act as a teacher model that can be distilled into smaller agents. According to Amazon, Nova 2 Pro is equal or better on 10 of 16 benchmarks versus Claude Sonnet 4.5, 8 of 16 versus GPT 5.1, 15 of 19 versus Gemini 2.5 Pro, and 8 of 18 versus Gemini 3 Pro Preview.

Nova 2 Sonic is a speech-to-speech model that unifies speech understanding and generation for real-time conversational AI. It supports English, French, German, Italian, Spanish, Portuguese, and Hindi, offers expressive voices with more natural turn-taking, and also exposes a one-million token context window so applications can maintain long-running voice sessions. Sonic adds cross-modal support so users can mix typed and spoken input in a single session and asynchronous tool calling so it can keep talking while tools run in the background.

Nova 2 Omni is described as a unified multimodal reasoning and generation model that can ingest text, images, video, and speech and generate both text and images. Amazon says it can handle up to 750,000 words along with hours of audio and long-form video, which lets it analyse large product catalogues, support documents, and media libraries in one go rather than coordinating several narrower models.

To give enterprises more control, AWS launched Nova Forge on 2 December 2025 as a program for building custom frontier models, called Novellas, using Nova checkpoints. Customers can start from pre-trained, mid-trained, or post-trained Nova 2 Lite and Nova 2 Pro checkpoints, mix their proprietary data with Nova curated datasets at each stage, and run training on Amazon SageMaker AI. Nova Forge also supports reinforcement learning in customer-defined environments and provides a responsible AI toolkit for safety and content controls. Early users include Booking.com, Cosine AI, Nimbus Therapeutics, Nomura Research Institute, OpenBabylon, Reddit, and Sony.

Nova Act, announced in the same About Amazon post, is a new AWS service for building and managing browser-based agents that carry out tasks in web user interfaces. Powered by a custom Nova 2 Lite model trained through large-scale reinforcement learning, Nova Act targets workflows like regression testing, CRM updates, and repetitive data entry. Amazon reports around 90 per cent reliability on early customer workloads and highlights case studies from Sola Systems, 1Password, Hertz, and Amazon’s own Leo team for satellite internet.

Alongside the models and services, AWS introduced what it calls frontier agents, a new class of long-running, autonomous agents that behave more like virtual team members than one-shot copilots. The first three are Kiro autonomous agent for software development, AWS Security Agent for secure by design engineering and penetration testing, and AWS DevOps Agent for incident response and operational improvement. These agents are described as autonomous, scalable, and capable of running for hours or days without human intervention, with AWS reporting that the DevOps Agent has achieved an internal root cause identification rate above 86 per cent. All three agents are available in preview.

Agent orchestration is handled by Amazon Bedrock AgentCore, which also received major enhancements on 2 December 2025. New Policy capabilities let teams define what tools and data an agent can use in natural language, with enforcement integrated at the gateway level so actions are checked in milliseconds. AgentCore Evaluations adds 13pre-builtt evaluators for dimensions such as correctness, tool selection accuracy, safety, goal success, and context relevance, while still allowing custom evaluators. A new AgentCore Memory feature introduces episodic memory so agents can store and reuse structured episodes that capture context, reasoning, actions, and outcomes across interactions.

Why it matters for learners, developers, and teams

The 2025 Nova announcements are less about another big model and more about turning agents into a structured platform. Instead of asking developers to wire up their own orchestration, memory, and evaluation layers around raw models, AWS now offers a stack that stretches from models to tools to opinionated, production-first agents. For teams already on AWS, that lowers the barrier to running long-lived agents that touch real infrastructure, code, and customer data.

For individual learners and early career developers, Nova 2 Lite and Nova 2 Sonic provide accessible entry points into building agents that can reason, call tools, and hold long contexts without managing multiple models. Being able to read the Nova documentation, call Nova 2 Lite from Bedrock, and integrate AgentCore policies and evaluations into a simple prototype is now a concrete learning target.

For experienced engineers, the existence of Nova Forge, Nova Act, and frontier agents pushes AI work into more structured roles. Skills that combine software engineering, distributed systems, cloud architecture, and MLOps will matter more than prompt engineering alone. Teams will need people who can design task graphs for Kiro, interpret evaluation dashboards, negotiate AgentCore policies with security and legal, and reason about when to use Nova Forge rather than simple fine-tuning.

From an organisational point of view, AWS is pressing the idea that AI agents should be governed like other critical systems. Policy, evaluation, and memory are the levers it offers for that governance. Anyone building or auditing such systems needs to understand how those levers work, not just how to call a model API.

Key numbers

Nova 2 Lite is equal to or better on 13 of 15 benchmarks versus Claude Haiku 4.5, 11 of 17 versus GPT 5 Mini, and 14 of 18 versus Gemini Flash 2.5 in its intelligence class.

Nova 2 Pro is equal to or better on 10 of 16 benchmarks versus Claude Sonnet 4.5, 8 of 16 versus GPT 5.1, 15 of 19 versus Gemini 2.5 Pro, and 8 of 18 versus Gemini 3 Pro Preview, according to Amazon.

Nova 2 Lite and Nova 2 Sonic both offer up to one million token context window to support long-running text and voice interactions.

Nova 2 Omni can process up to 750,000 words along with hours of audio and long videos in a single pass.

Nova Act achieves around 90 per cent reliability on early browser-based workflows and has been used to speed Hertz software delivery by a factor of five.

Trainium3 UltraServers pack 144 Trainium3 chips and deliver up to 4.4 times more compute performance and 4 times better energy efficiency compared with the previous Trainium generation, underpinning the Nova and agent offerings.

AWS reports that its DevOps Agent has an internal root cause identification rate above 86 per cent on thousands of incidents.

Context

Amazon started rolling out the Nova family earlier in 2025 as a suite of multimodal foundation models exposed through Amazon Bedrock. Nova was already positioned for text, image, video, speech, API calling, and agentic AI, but the 2 December 2025 announcements formalise Nova 2 as a four-model family tuned for reasoning, real-time speech, and multimodal generation, plus an ecosystem of embeddings and tools built around agent use cases.

The Nova 2 lineup sits in a crowded market. Amazon’s own benchmark tables compare Nova 2 Lite to models such as Claude 4.5 Haiku, GPT 5 Mini, and Gemini Flash 2.5, and Nova 2 Pro to Claude Sonnet 4.5, GPT 5, GPT 5.1, Gemini 2.5 Pro, and Gemini 3 Pro Preview. External coverage from Wired notes that Amazon claims Nova 2 Pro matches or exceeds those competitor models on a range of benchmarks, especially for agentic tasks such as tool use and multi-step reasoning, although adoption still lags leading APIs.

What sets Nova 2 apart is less the raw benchmark positioning and more the integration story. Nova Forge exposes checkpoints and recipes that are normally reserved for model providers, letting customers build domain-specific frontier models without starting from scratch. Nova Act applies a similar idea to UI automation, providing a managed way to run agents in browsers. Frontier agents like Kiro, AWS Security Agent, and AWS DevOps Agent show how Amazon itself is using those building blocks to augment software teams.

Reuters coverage of re: Invent 2025 frames these moves as part of a wider effort to keep AWS competitive against Nvidia, Microsoft, and Google by combining its own Trainium chips, Nova models, and services such as Nova Forge into an integrated AI infrastructure story. For cloud customers, that means the question is less which single model to use and more which stack to bet on for the next few years, agent-heavy workloads.

What is next

In the near term, the most practical experiments will revolve around Nova 2 Lite and Nova 2 Sonic, since they are generally available in Amazon Bedrock and have clear entry points in the console and SDKs. Teams can start by swapping these models into existing chatbots, support tools, or contact centre pilots that already run on Bedrock, then layering in web grounding, code execution, and extended thinking where it makes sense.

As Nova Forge programs expand, more organisations will treat their Novellas as strategic assets rather than one more fine-tuned model. That will require better data governance, evaluation pipelines, and red teaming practices, since a misconfigured frontier-grade internal model can rapidly propagate mistakes through fleets of agents. Understanding AgentCore policies and evaluations early will make these deployments smoother.

Frontier agents and Nova Act are likely to evolve based on early customer feedback. Developers should watch how Amazon itself uses these tools inside large engineering organisations, because those patterns will influence how hiring managers assess experience with agentic AI. Knowing how to integrate agents with CI and CD, observability, and incident management tooling will be at least as important as writing good prompts.

For learners and career changers, the roadmap is becoming clearer. Build solid foundations in Python, data structures, cloud basics, and security. Add practical experience with Amazon Bedrock, Nova 2, and AgentCore, then start experimenting with small agents that solve real problems in your own projects or workplaces. That combination of fundamentals and hands-on cloud AI work is what hiring managers will look for as agent-heavy systems move into production.

How to go deeper

Explore generative modelling, evaluation techniques, and agent design patterns with Code Labs Academy’s Data Science and AI bootcamp. The curriculum covers Python, SQL, and modern machine learning workflows that map well to Nova 2 and Bedrock. See the Data Science and AI course page at Code Labs Academy.

If you want to build the APIs and front ends that Nova-backed agents will plug into, review the Web Development bootcamp, which covers JavaScript, back-end basics, and deployment patterns that mirror real-world stacks.