July 28, 2026, (Inside AI) — Amazon is winding down most of its in-house Nova AI models, including the high-end Premier, Omni, Reel video-generation, and Canvas image-generation models, according to a Business Insider report citing people familiar with the matter. The company is instead shifting resources toward a new frontier-model effort led by researcher Pieter Abbeel, with frontier research becoming a top priority this year.
The move marks a significant pivot for Amazon's artificial general intelligence (AGI) strategy, just days after the company trimmed jobs in its AGI group. A spokesperson said last week that Amazon was sharpening its focus on "the initiatives that matter most for customers." The new flagship model is expected to debut at Amazon's annual re:Invent conference later this year, possibly under the Nova brand.
Amazon has struggled to generate buzz around its Nova models compared to rivals like OpenAI, Anthropic, and Google. Instead, the company has doubled down on providing cloud and AI infrastructure through AWS. This strategic overhaul suggests Amazon is conceding that its current in-house models cannot compete on performance or mindshare, while betting on a concentrated push into frontier AI.
The deprecation of most Nova models is a stark admission. The Nova family was launched with fanfare, but industry analysts note that Amazon's models never achieved the benchmark scores or developer adoption of GPT-4 or Claude. By contrast, Amazon's Bedrock platform has thrived by offering third-party models, a strategy that reduces reliance on in-house AI. This realignment may reflect a broader trend: hyperscalers are finding it more profitable to host and fine-tune external models than to build their own from scratch.
Pieter Abbeel, a renowned roboticist and AI researcher, joined Amazon in 2024 after co-founding Covariant, a robotics AI startup. His appointment signals a focus on embodied AI and reasoning capabilities that could differentiate Amazon's next model. Frontier models typically require massive compute and novel architectures, and Amazon's cloud infrastructure gives it a unique advantage in scaling such efforts.
However, the AGI group has seen turmoil. Amazon consolidated its AGI work under cloud executive Peter DeSantis in December 2025, and several top AGI executives have departed over the past year. The job cuts last week underscore the pressure to deliver results. A source familiar with the matter told Business Insider that the new model is a "top priority," but Amazon has not disclosed technical details or benchmarks.
Amazon is not abandoning Nova entirely. The company will continue to offer Nova 2 Lite and Nova 2 Sonic models, as well as Nova Forge, a service that lets customers build custom AI models. This suggests a tiered approach: lightweight, cost-effective models for broad use, and a frontier model for cutting-edge tasks. Yet the decision to kill Premier and Omni indicates that the middle ground is not viable against competitors.
Contextually, this pivot mirrors Microsoft's early reliance on OpenAI and Google's internal struggles with model fragmentation before the Gemini unification. Amazon's challenge is that frontier models are expensive and risky; even well-funded labs like Character.AI have struggled. But Amazon's deep pockets and customer base could allow it to iterate faster than startups.
What's missing from the report is any mention of safety protocols or alignment research for the frontier model. Given Abbeel's background in robotics, the model may emphasize real-world interaction, raising questions about responsible deployment. Amazon has not published a detailed safety framework for Nova, and this opacity could draw regulatory scrutiny as governments tighten AI rules.
The re:Invent conference in Las Vegas will be a critical moment. If Amazon unveils a model that matches or exceeds GPT-5 or Claude 4, it could reshape the AI landscape. If not, the company may further retreat to its infrastructure stronghold. For now, the winding down of Nova models is a clear signal: Amazon is betting big on a single, high-stakes AI moonshot.
