The Seattle giant announces new AI services focused on business needs as well as a comprehensive update to its Nova family of models.
AI applied to the real needs of businesses. This is the common thread that AWS is clearly highlighting for this new edition of re:Invent, the annual high mass of the cloud which is held from December 1 to 5 in Las Vegas. The group is no longer content with stacking models: it seeks to equip organizations to industrialize AI and agents. On the program this Tuesday, December 2: a major update of its LLM Nova family, a platform designed to build your own model with business data, and a new generation of agents capable of operating in (almost) complete autonomy on complex business tasks. The JDN, present on site, reveals the main new features.
Amazon Nova 2: a quality/price oriented LLM family
This is the main announcement of the day. Amazon unveils a major update to its Nova family of models, launched just a year ago. The company claims to have worked specifically on the performance/price ratio in order to offer competitive models, capable of producing cutting-edge results at an attractive cost. Four models are presented this Tuesday: Nova 2 Lite, Nova 2 Pro, Nova 2 Sonic and Nova 2 Omni. Each targets specific use cases identified by Amazon.
Nova 2 Lite is the fastest model, designed to provide very low latency responses on simple tasks. It is positioned as an equivalent to Gemini 2.5 Flash, with minimal reasoning. It can be used for basic stains.
Nova 2 Pro constitutes the cornerstone of this new range: an advanced reasoning model designed for the most complex tasks. Amazon claims that it meets or exceeds Claude Sonnet 4.5, GPT-5.1 and Gemini 3 on the majority of benchmarks, although these results have not yet been published at the time of writing this article. Nova 2 Pro is presented as the most intelligent model ever developed by Amazon, reserved for uses that require analysis and reasoning.
Nova Sonic is the speech-to-speech version of the range: useful for voice interaction, but probably limited to basic scenarios, far from the capabilities of Lite and Pro.
Nova 2 Omnifinally, completes the family with a natively multimodal model, capable of understanding audio, video (and images) and generating images. It remains to evaluate its actual performance: there is currently no guarantee that it already rivals the capabilities of Nano Banana.
The exact pricing of the models should be clarified quickly after the launch this Tuesday.
Amazon Nova Forge: to create your own LLM with company data
This is another major novelty. AWS announces a platform entirely dedicated to the creation of tailor-made LLMs, by combining company data with datasets selected by Amazon. The positioning is clear: target organizations that do not have the budget to train a model from scratch, and those that want to avoid the problems of catastrophic forgetting inherent in fine-tuning.
Based on the weights of the Nova models, Forge will train an “in-house” model from the company’s internal data, while supplementing it with datasets pre-trained by Amazon. The idea is to obtain a model truly adapted to business needs, without excessive complexity. Once the model is finalized, it can be deployed immediately to Bedrock for inference.
Frontier Agents: new agents focused on almost complete autonomy
It is another product focused on applied AI, with very concrete added value for businesses. AWS unveils a new range of agents designed to operate with a high level of autonomy and native scalability. The goal: agents capable of operating alone for several days, with a minimum of human interaction. This Tuesday, three frontier agents are presented: Kiro Autonomous Agent, AWS Security Agent and AWS DevOps Agent.
Kiro Autonomous Agent acts as a virtual developer capable of maintaining context, chaining tasks and pushing pull requests without supervision. It manages the background work of a developer: bugs, tests, minor changes and integrates directly with GitHub, Jira or Slack. It allows you to automate basic tasks, well beyond a simple code agent.
AWS Security Agent is designed as a security engineer embedded directly in the development cycle. It analyzes the code, audits pull requests according to the company’s internal policies and detects possible flaws. It can even run full pentests on demand, with fixes generated automatically. The goal is to have a system capable of analyzing everything continuously and being proactive.
Finally, AWS DevOps Agent functions as a senior DevOps specialist. It reacts instantly to incidents, correlating logs, telemetry and the latest deployments to find the cause of a bug or error. It then recommends to the human the corrective actions to be carried out.
New Mistral models available on Bedrock at launch
Amazon is also boosting Bedrock, its AI platform most popular with developers. The group is announcing this Tuesday its largest extension of model catalogs since the launch of the service, with the arrival of 18 new open source models from the main market players. Among them, two major new features from Mistral AI: Mistral Large 3, an advanced model optimized for long contexts, multimodal and precise following of instructions and Ministral 3, a more compact model designed to be fast.
With re:Invent, AWS is rolling out an edition once again focused on AI, but with a positioning that is becoming clearer: the emphasis is on business use cases rather than technological demonstration. The speech, like the products announced, seeks to anchor AI in operations. It remains to be seen whether this very business-oriented strategy will allow Amazon to stand out in a landscape where all players are now demanding more useful and more integrated AI.
