Amazon Nova Premier 1.0: A Teacher Model at End of Life

If you are reading about Nova Premier because you are about to build on it, stop and read the next section first. Premier was Amazon's most capable first-generation Nova model, but in 2026 it was moved onto AWS's retirement schedule, and the announced end-of-life date on Bedrock has now passed.

The model is still worth understanding. It was designed for a job most models are not, teaching smaller models, and that idea outlives it. But the practical content of this guide is mostly about what to use instead.

Status first: Premier is at end of life

According to AWS's Bedrock lifecycle schedule, as reported in July and August 2026, Nova Premier entered the Legacy state on 13 March 2026 with an end-of-life date of 14 September 2026. Press coverage tied this to Amazon scaling back several Nova models to concentrate on a new frontier-model effort, while Nova 2 Lite was reported as continuing.

Bedrock's lifecycle rules spell out what that means. During the Legacy period new customers cannot adopt the model, and existing customers can lose access after 15 days of inactivity. After the end-of-life date, requests fail in the Regions covered unless there is a private arrangement with the provider. Migration does not happen automatically.

Third-party catalogues can lag behind these changes; at the time of writing Premier still appears in OpenRouter's listing at $2.50 per million input tokens and $12.50 per million output. Whether any given route still serves it is something to check with a live request, not assume. On Bedrock, the modelLifecycle field returned by ListFoundationModels is the authoritative answer for your Region.

What Premier was built to do

Amazon first announced Premier in December 2024 alongside Micro, Lite and Pro, promising it for early 2025. It reached general availability on 30 April 2025. The launch post called it Amazon's most capable multimodal model for complex tasks and, unusually, its most capable teacher model.

The specification was built for large inputs. AWS documented a 1M-token context window, enough by its own description for documents over 400 pages or about 90 minutes of video, with text, image and video inputs and text output. Access at launch was through cross-Region inference in US East (N. Virginia and Ohio) and US West (Oregon).

The teacher role is the distinctive part. With Bedrock Model Distillation, Premier generated training data to produce customised versions of Nova Pro, Lite and Micro. AWS's example was a Nova Pro distilled for tool selection and API calling, which it reported as 20% more accurate on API invocations than the base Pro. The Nova Pro guide covers that student side.

The benchmark sheet, read honestly

AWS published these scores at launch:

• MMLU: 87.4%
• Math500: 82.0%
• CharXiv (chart understanding): 84.6%
• SimpleQA with RAG: 86.3%
• BFCL (function calling): 63.7%
• SWE-bench Verified: 42.4%

The strong numbers are on knowledge, charts and retrieval-backed question answering. The weak ones are on function calling and agentic coding. Contemporary coverage noted it was not a reasoning model and that it trailed the leading models of spring 2025 on coding. A 42.4% on SWE-bench Verified was a long way behind the frontier then and is further behind now.

These are vendor-run numbers, with the usual caveats. Vendor-claimed vs independent benchmarks covers how to discount them. The shape of the profile is more informative than any single figure: Premier was a long-context reader, not an agent.

Moving off it

AWS's own migration guidance points Premier users to Nova 2 Lite, specifically for agentic and tool-use workloads, and recommends evaluating it across reasoning effort levels. AWS claims Nova 2 Lite beats Premier on multi-step problem solving at about seven times lower cost.

Nova 2 Lite keeps the 1M window and the same input types, adds optional reasoning and much longer output, and lists at $0.30 in and $2.50 out on OpenRouter. On paper it is a straightforward replacement. In practice, test the parts of your workload that used Premier's strengths:

1. Long-document answers. Run your largest real inputs and check whether details from the middle of the document survive.
2. Chart and figure reading. Premier's CharXiv score was its standout; verify Nova 2 Lite on your own charts.
3. Distillation pipelines. If Premier generated your training data, decide whether to regenerate with a new teacher or freeze the existing dataset.
4. Prompt sensitivity. Prompts tuned for a non-reasoning model can behave differently once reasoning is enabled.

If Nova 2 Lite falls short, Bedrock's third-party catalogue is the other route. Model deprecation and migration has a general plan for exactly this situation.

What to keep from the teacher idea

Premier's lasting lesson is the workflow it was marketed for. A large, expensive model is used briefly to generate examples for one narrow task, and a small model is tuned on them and run at volume. The expensive model's cost is paid once, not on every request.

That pattern does not depend on Premier. Any capable model whose terms allow it can play teacher, and the student you trained is yours to keep running. The takeaway: do not start new work on Premier, migrate what remains before access disappears, and keep the distilled models you built with it if they still pass your tests.