Amazon spent the past 20 months telling developers that its homegrown Nova models were a serious answer to OpenAI, Anthropic and Google. In late July 2026, it quietly conceded that most of them were not. According to a Business Insider report corroborated by multiple outlets, Amazon has begun winding down active development on the bulk of its flagship Nova lineup and is pouring the freed-up engineers and compute into a single, concentrated bet: one frontier-scale foundation model built by a brand-new internal group.
The retrenchment is broad. Nova Premier, Amazon's top large language model, along with the multimodal Nova Omni, the Reel video generator and the Canvas image generator, are being moved into what employees described as "keep the lights on" mode — maintained and supported for existing customers, but no longer a priority for new engineering investment. The KTLO designation, familiar to anyone who has watched a product's roadmap freeze, is about as clear a signal of deprioritization as a company gives without formally sunsetting a product.
Not everything is going away. Amazon is retaining Nova 2 Lite, its efficient workhorse model; Nova 2 Sonic for speech; Nova Forge, the service that lets customers build customized models on Amazon's technology; and Nova Act, its agentic browser-automation model. What Amazon is abandoning is the idea of fielding a full parallel family of specialized systems for text, image, audio and video. In its place: Frontier Model Research, or FMR, a new team reportedly led by Pieter Abbeel, the UC Berkeley robotics professor who joined Amazon through its acquisition of the startup Covariant. The mandate is narrow and audacious — build one model that developers actively prefer over Claude, GPT and Gemini.
Amazon did not frame the move as a retreat. "AI models remain one of the most important things we're working on, and that hasn't changed," an Amazon spokesperson told Business Insider. The company added that "we're continuing to invest in and support the Nova models our customers rely on today," while "also investing in the next generation of frontier model research." In other words: less breadth, more depth.
Why building a frontier model is so hard
The strategic logic is sound, and it is also an admission. Nova, launched with fanfare at re:Invent in December 2024, never dislodged the incumbents in benchmark rankings or developer mindshare. Spreading a finite pool of elite researchers and scarce accelerator capacity across four or five distinct model families is a recipe for finishing second in all of them. Concentrating that firepower on a single frontier system is how OpenAI and Anthropic operate.
But frontier modeling is punishing precisely because the leaders keep moving. Reaching parity with GPT or Claude is not a fixed target; it is a race against labs that ship new generations every few months and have spent years compounding advantages in data pipelines, post-training and reinforcement learning from human feedback. Talent is the other constraint. The reorganization has come with real costs: Amazon has closed its AGI Lab and reduced headcount, and since December its AI efforts have been consolidated under Peter DeSantis, the AWS infrastructure chief, with Abbeel reporting into that structure. Reshuffling org charts and shedding staff is not obviously the fastest path to a breakthrough.
The Anthropic paradox
The deeper tension sits one layer down the stack. Amazon has poured a multibillion-dollar investment into Anthropic, the maker of Claude, and Anthropic has committed to spending more than $100 billion on AWS over roughly a decade while securing gigawatts of Amazon's custom Trainium compute. AWS also hosts models from OpenAI. Amazon, in short, is the landlord for the very companies whose products have outrun Nova.
That arrangement is enormously lucrative — AWS is growing more than 30% year over year, and Amazon has signaled around $200 billion in planned capital expenditure for 2026, much of it AI infrastructure. It also raises an awkward question: if Claude runs beautifully on AWS and customers are happy, how hard will Amazon really push a competing in-house frontier model, and how much will developers trust it over the alternatives sitting right beside it in the same console? Amazon's most reliable AI profits today come from renting infrastructure to its rivals, not from beating them at modeling.
What to watch
The proof point is re:Invent this autumn. Amazon is expected to unveil its first FMR-built flagship there — possibly still under the Nova brand — and that debut will answer the questions this reset has raised. Does the model land within striking distance of the frontier, or a generation behind? Does Amazon price and position it to genuinely compete with Claude on its own cloud, or treat it as a hedge? And does the developer community, which shrugged at the original Nova, give a consolidated second effort a real look?
For now, Amazon has made the harder, more honest choice: stop pretending it can win everywhere and bet on winning once. Whether Abbeel's team can turn that focus into a model the market actually reaches for is the wager that will define Amazon's AI standing heading into 2027.
"AI models remain one of the most important things we're working on, and that hasn't changed."— Amazon spokesperson, Company statement to Business Insider