AWS dropped 50+ announcements at re:Invent. Here’s an attempt to capture all the main ones…
First up
Nova 2 = finally competitive. They released four new models: Lite (adjustable reasoning), Pro (the smartest, serves as “teacher” for distillation), Sonic (for real-time speech, 1 million token context), and Omni (first to reason AND generate images simultaneously). Try them out here.
In addition to that, they also released…
- Nova Forge, which let’s you build custom models for $100K/year. The catch = models are locked to Bedrock.
- Nova Act, which = 90% reliable browser automation.
- AWS Transform, which = 5x faster code modernization.
- Plus support for 18 new open-weight models, including new Mistral Large 3 (more on that below), OpenAI gpt-oss, Qwen3, DeepSeek-V3.1, etc.
Next up
New Frontier Agents, a.k.a new AI coworkers. These are three autonomous agents work hours/days without intervention:
- Kiro: virtual developer, learns your patterns, and Amazon’s official internal tool (although…).
- Security Agent: Does code reviews, pen testing, and files tickets vs. auto-fixing.
- DevOps Agent: Commonwealth Bank used it to solve incidents in 15 minutes vs. hours.
They also released Bedrock AgentCore, which = the production platform. It supports any agent framework (CrewAI, LangGraph, OpenAI SDK, etc) and now includes Memory, Browser Tool, and Observability.
And, of course, there’s the infrastructure:
- New Trainium3 UltraServers, which = 144 chips, 362 petaflops, 4.4x performance. Customers cutting training costs 50%. Whoa dude.
- Trainium4 (coming in 2026), which will have 6x performance AND support NVIDIA NVLink Fusion.
- Plus AI Factories, which = dedicated AWS infrastructure in your data center.
Why this matters
In case you thought Amazon wasn’t a serious contender in AI, this launch is meant to remind you it is very serious about AI. While the new models don’t seem earth-shattering, the Nova Forge is unique. Amazon says no other company lets you train your own foundation models the way they do.
The problem is you can’t take them off Amazon because they’re not giving you the actual weights (which = the numbers the AI uses to know how to respond). So TBD if customers actually find value in this type of managed service. I guess you’re doing the same thing more or less when you use OpenAI’s closed models, except in this case, you’re training your own model…
Editor’s note: This content originally ran in the newsletter of our sister publication, The Neuron. To read more from The Neuron, sign up for its newsletter here.