Amazon Nova Premier

What it is

Amazon Nova Premier is the most capable model in the first generation of Amazon's Nova family of foundation models. It became generally available in Amazon Bedrock, Amazon's service for building applications on top of foundation models, on 30 April 2025.^[1] The model takes text, images and video as input, though not audio,^[1] and it can work with up to 1 million tokens of context at a time.^[1] It comes with built-in safety and content-moderation controls.^[1]

Where it came from

Amazon first announced Nova Premier on 3 December 2024, when the model was still in training and was slated to arrive in early 2025.^[2] The April 2025 launch was the point at which developers could actually call the model through Bedrock rather than only hear that it was coming.

What Amazon says it can do

Amazon pitched the model at complex tasks that require a deep understanding of context, multistep planning, and precise execution across multiple tools and data sources.^[1] In Amazon's own evaluation, Nova Premier was the most capable Nova model across 17 benchmarks, and it was equal to or better than models in the same tier on about half of them.^[1] AWS described it as comparable to the best non-reasoning models in the industry, and equal or better than those on roughly half of the benchmarks in its tier.^[1] AWS also said it was the fastest and most cost-effective model in Amazon Bedrock for its intelligence tier.^[1]

Outside estimates of cost came from Slack, which said that in its tests the model was "nearly half the cost compared to other leading models".^[1] Such figures depend on the workload being run, so they describe a particular set of tests rather than a general price list.

A teacher model as well as a product

One role set Nova Premier apart from the rest of the family: Amazon positioned it as a teacher model for distillation into the smaller Nova Pro, Lite and Micro models.^[1] Distillation means training a smaller model on the outputs of a larger one, so that the cheaper model picks up part of the larger model's behaviour without carrying its full cost. Amazon gave one concrete measurement of the effect: a Nova Pro distilled from Nova Premier for tool selection was 20% more accurate on API invocations than the base model.^[1]

That framing matters for how the model is used. Nova Premier is both a model that developers can call directly for demanding work and a source of training signal for the models most applications would actually deploy.

Where it runs

At launch the model was offered in three regions — US East (N. Virginia), US East (Ohio) and US West (Oregon) — through cross-Region inference, under the model ID us.amazon.nova-premier-v1:0.^[1] Cross-Region inference means a request can be routed to capacity in another region within the same group, which helps when demand in one location is high.

Where it sits in the family

Nova Premier is a first-generation model, and Amazon placed it at the top of that generation on capability while describing it as belonging to the non-reasoning tier of the market rather than the reasoning tier.^[1] Its distinguishing features are the combination of a very long context window, mixed text, image and video input, and the additional job of supplying training signal to smaller models in the same family. The benchmarks Amazon reported are its own comparisons against models in the same tier, and the claims of speed and cost-effectiveness are likewise about where the model sits within Amazon Bedrock for that tier.^[1]