Amazon launches Nova Premier, its ‘most capable’ AI model yet

Amazon Web Services (AWS) has launched Nova Premier, its most advanced AI model to date, via Amazon Bedrock. Designed for enterprise use, the model targets complex, multi-step workflows and supports model distillation, enabling smaller models to inherit its capabilities with improved efficiency and reduced cost.

Nova Premier can handle text, image, and long-form video inputs with a one-million-token context window (equivalent to around 750,000 words) and will support over 200 languages, according to an AWS blog. Nova Premier’s applications would span financial analysis, software automation, and agentic tasks involving orchestration across tools and data layers, AWS said.

According to Deepika Giri, head of research – BD & AI at IDC Asia/Pacific, Nova Premier distinguishes itself by applying LLMs to niche Agentic AI scenarios where performance and cost-efficiency are critical. “Its multimodal capabilities also expand its relevance across a wide range of enterprise use cases,” she said.

While AWS describes it as its “most capable model,” Nova Premier trails key rivals in some third-party benchmarks. It lags behind Google’s Gemini 2.5 Pro in coding (SWE-Bench Verified) and scores lower on math and science evaluations such as GPQA Diamond and AIME 2025. However, Amazon’s internal testing shows Nova Premier performs strongly in knowledge retrieval and visual reasoning, with scores of 86.3 on SimpleQA and 87.4 on MMMU.

Pricing is in line with industry standards — $2.50 per million input tokens and $12.50 for output tokens — comparable to Google’s Gemini 2.5 Pro.

Distillation capabilities

A key feature of Nova Premier is its support for model distillation within Bedrock. This would allow enterprises to generate synthetic data from Premier and fine-tune smaller models such as Nova Pro, Lite, and Micro for targeted applications.

According to AWS, a distilled Nova Pro model increased API invocation accuracy by 20% while delivering similar output at lower cost and latency. The process eliminates the need for labeled training data and is positioned for edge deployments and use cases with resource constraints.

“Distillation enables customers to create smaller, more efficient models for specific tasks,” AWS noted. This sets it apart from other approaches. OpenAI’s GPT-4o-mini, for instance, leans on fine-tuning, while Anthropic’s Claude prioritizes text optimization.

Amandeep Singh, principal analyst at the QKS Group, said the launch is a strategic shift for AWS. “Nova Premier signals AWS’s move from a neutral model host to asserting foundational control in the GenAI value chain. This isn’t about building the biggest model, it’s about owning the orchestration, pricing, and architecture.” He added that combining proprietary models with Bedrock’s flexible interface strengthens AWS’s appeal as a sustainable AI stack for enterprises.

Currently, Nova Premier is available only to approved Bedrock users.