Amazon kills most Nova AI models, shifts team to Frontier group under Pieter Abbeel

Amazon is gutting its in-house Nova AI lineup. The flagship Nova Premier and Omni models, the Reel video model, and the Canvas image generator are all effectively on life support — still available for existing customers but no longer in active development. That’s a fast retrenchment for a product family unveiled with fanfare only in December 2024. The pullback comes alongside layoffs in the AGI division and the quiet closure of the AGI Lab Amazon set up after acquiring Adept.

But this isn’t Amazon waving a white flag. It’s a hard pivot. Resources are now funneling into a new Frontier Model Research group led by Pieter Abbeel, the robotics and reinforcement learning researcher who arrived via the Covariant acquisition. A new foundation model is reportedly on the docket for the re:Invent conference this fall. The message is clear: Amazon is done trying to compete across a broad assortment of me-too generative models and is instead betting on a single, high-stakes moonshot.

A spokesperson told Business Insider that AI models remain one of Amazon’s most important projects, and the company’s checkbook backs that up. It’s one of Anthropic’s biggest backers and has deep ties to OpenAI, with combined investments that could eventually top $100 billion. For a company that plays the long game, pouring that kind of money into rivals while axing its own models is a sobering admission: the first wave of Nova wasn’t good enough, and building frontier-quality models from scratch is harder than even a trillion-dollar company anticipated.

What’s genuinely interesting here is the who, not just the what. Putting Abbeel in charge signals a shift toward models that might reason and act in the physical world — a sharp departure from the text-and-image race. If Amazon can’t beat Anthropic at chatbots, maybe it can build something OpenAI and others aren’t even working on yet. That’s the bet. Whether it pays off before the market moves on is anyone’s guess.