At last December's AWS re:Invent, Amazon did something interesting: instead of just releasing new models, it bundled "models + training platform + automation services" into a single package.
The Nova 2 family launched with four models at once, alongside Nova Forge (letting enterprises train their own model variants) and Nova Act (browser-level AI automation). Individually, none is the brightest star, but together, the three-piece set makes Amazon's strategy clearer.
Four Models, Each with a Role
Nova 2 Lite is positioned for cost-effectiveness, handling text, images, and video, with support for "adjustable thinking depth" — meaning you can set how much compute to spend on reasoning before inference. On benchmarks, the company claims it is no worse than Claude Haiku 4.5 in 13 out of 15 tests, compared to GPT-5 Mini at 11/17, and Gemini Flash 2.5 at 14/18.
Nova 2 Pro is the flagship, specializing in multi-document analysis, video reasoning, and complex math. It won 10 out of 16 tests against Claude Sonnet 4.5, and 15 out of 19 against Gemini 2.5 Pro — a strong showing on Google's own turf.
Nova 2 Sonic is a speech-to-speech model with a 1M token context, multilingual, targeting real-time conversational scenarios, and can directly integrate with Amazon Connect and telecom operators.
Nova 2 Omni is the most ambitious attempt: unified multimodal input and output, handling text, images, video, and speech input, with text and image output, all in one model. Amazon calls it the "industry's first" architecture of its kind — capable of processing 750,000 words, hours of audio, and long videos simultaneously.
Both core models come with built-in web grounding (real-time internet retrieval) and code execution tools, requiring no additional configuration.
Nova Forge: Let Enterprises Train Their Own Models
This is the most interesting part of the package. Nova Forge is not fine-tuning; it is "open training" that incorporates enterprise private data from the training stage — the resulting model variants are called Novellas.
Beyond basic training, Nova Forge also offers: reinforcement learning training environment (gym), synthetic data distillation, and a responsible AI toolkit.
Early users include Reddit, Booking.com, Sony, Cosine AI, and Nimbus Therapeutics. Reddit CTO Chris Slowe said they replaced multiple different models with a Novella, using a more accurate solution for content moderation, with clear efficiency gains.
From a business logic perspective, Nova Forge gives large customers a reason: don't go to Anthropic or OpenAI's fine-tune API; come to AWS to train your own version from the ground up. This differentiator has no direct comparable product from other AI vendors.
Nova Act: 90% Reliability in Browser Automation
Nova Act is a service for building browser-controlling agents, powered by a specially trained Nova 2 Lite variant running reinforcement learning.
It is similar to Anthropic's Computer Use and OpenAI's Operator, but Nova Act emphasizes engineering reliability: early customers report achieving 90% reliability in real-world workflows. If this number holds, it is close to the threshold for production deployment — many "computer use" products in the industry have actual reliability in the 60-70% range, a significant gap.
Several case studies provide clear numbers:
| Customer | Result |
|---|---|
| Hertz | 5x faster software delivery, QA bottleneck eliminated |
| Sola Systems | Hundreds of thousands of workflows automated monthly |
| 1Password | Reduced manual steps in login processes |
| Amazon Leo (internal) | Testing reduced from weeks to minutes |
Nova Act includes a no-code playground for writing instructions in natural language, and a VS Code plugin for deploying code versions.
What Amazon Is Betting On
Putting the three-piece set together, Amazon's strategy is: lock enterprises into the Bedrock ecosystem, covering everything from general models to custom training to automated execution in one place.
The model capabilities of Nova 2 may not be the absolute frontier, but for enterprises already running workloads on AWS, the switching cost is extremely low. More importantly, Nova Forge's custom training capability is a true differentiator — the Novella you train is yours, and cannot run directly on a competitor's platform.
Interestingly, Amazon itself is using Nova 2: the internal test project Amazon Leo uses Nova Act to compress software testing time from weeks to minutes, essentially an official endorsement.
Developers can register for a free tier at nova.amazon.com/dev. The full suite is deployed through Amazon Bedrock.
Source: Amazon introduces new frontier Nova models, a pioneering Nova Forge service for organizations to build their own models, and Nova Act for building agents (aboutamazon.com, CocoLoop, 2025-12-02)