HomeAI models
Amazon: Nova Premier 1.0
Amazon
•
Nova Premier
•
Released 
April 2025

Amazon
Nova Premier

Amazon's most capable Nova model, built for million-token context, multimodal input, and distilling smaller custom Nova variants.

Modality:
Text
Image
PDF
model ID
amazon/nova-premier-v1
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed

62.66
tok/s

Intelligence Index

12.7
/ 100

Context Window

1
tokens

Input price

15
Anytoken

Output price

75
Anytoken
Nova Premier 1.0: Million-Token Context and Distillation Teacher for the Amazon Nova Family Nova Premier 1.0 is Amazon's most capable Nova understanding model, available in Amazon Bedrock. It accepts text, images, and video as input and returns text, paired with a one-million-token context window for long documents, large codebases, and long video. Positioned as the family flagship, it also serves as the primary teacher model for Bedrock Model Distillation, producing cheaper, faster custom variants of Nova Pro, Lite, and Micro. Its strongest fit is long-context understanding and multistep agentic workflows where a very large input window matters more than top-tier raw intelligence. Start building with Nova Premier 1.0 via the AnyAPI.ai API.

Performance

Where Nova Premier's Million-Token Window Earns Its Place

Nova Premier is strongest at long-context understanding: single-prompt analysis of large codebases, long documents, and up to roughly 90 minutes of video. Amazon reports 87.4% on MMLU and 82.0% on Math500, and documents that the model excels at code understanding and question answering over long documents. That combination matters when your bottleneck is fitting an entire corpus into one request rather than winning frontier reasoning benchmarks. In production, this makes Premier a practical choice for whole-repository or whole-document tasks, though Amazon notes accuracy can decline slightly as context grows, so input placement and structure affect results.

Benchmarks

Nova Premier Benchmarks: Capable, But Not Frontier

Independent testing tells a consistent story. Artificial Analysis places Nova Premier around 13 on its Intelligence Index, below the non-reasoning average near 18, and measures output near 34 tokens per second with a time to first token of about 2.91 seconds—slower than many similarly priced non-reasoning models. On coding, Amazon reports 42.4% on SWE-bench Verified, ahead of some older frontier models but behind leading Claude Sonnet versions. Premier also trails top competitors on graduate-level reasoning (GPQA Diamond) and competition math (AIME 2025). Treat it as a capable long-context model, not an intelligence leader.

Output Speed

62.66
tok/s

Intelligence Index

12.7
/ 100

MMLU

Broad world knowledge and problem-solving
73
%

GPQA

PhD-level scientific reasoning across physics, biology, chemistry.
57
%

HLE

Adherence to multi-step structured instructions.
4
%

LiveCodeBench

Tool-calling reliability in long agentic loops.
32
%

Technical Specifications

What the model supports

Nova Premier accepts text, image, and video input and produces text output only—there is no audio input or image/video generation. The defining specification is the one-million-token context window, which Amazon describes as roughly 1M text tokens, 500 images, or 90 minutes of video. Maximum output is comparatively modest per the Bedrock model card, so Premier is built for large inputs and concise-to-moderate outputs, not extremely long generation. It runs in Amazon Bedrock via the Converse and Invoke APIs with tool calling, structured JSON-schema outputs, streaming, and prompt caching, and supports over 200 languages.
Verified Specifications — 
Nova Premier
Input modalities
Text
Image
PDF
output modalities
Text
Context window
1
 tokens
Maximum output tokens
32000
Reasoning
No
Knowledge cutoff
April 2025
Pricing (standard)
15
 AnyTokens in
 / 
75
 AnyTokens out

Limitations & Trade-offs

Where Nova Premier falls short

1
Below-average raw intelligence. Independent testing (Artificial Analysis) places Nova Premier near the bottom of the non-reasoning tier on its Intelligence Index, below competitors at similar price points. Amazon itself describes Premier as equal or better on roughly half of its evaluated benchmarks against same-tier models. For tasks that demand strong reasoning, graduate-level science (GPQA Diamond), or competition math (AIME 2025), leading Claude, GPT, or Gemini models are better choices.
2
Not a reasoning model. Amazon's launch materials and independent analysis describe Nova Premier as a direct-response model without an extended chain-of-thought mode; it gives answers without a separately controllable thinking budget. Workloads that benefit from explicit step-by-step deliberation—hard math, complex planning, or difficult debugging—may be better served by dedicated reasoning models. Note the Bedrock model card lists reasoning as supported, which conflicts with the launch positioning, so verify behavior for your use case.
3
Slow and comparatively expensive. Independent measurements put output near 34 tokens per second and time to first token around 2.91 seconds—slower than many similarly priced non-reasoning models. Premier also sits in a higher-cost tier than smaller Nova models. For latency-sensitive or high-volume generation, Nova Pro, Nova Lite, or a distilled custom variant will be faster and more cost-effective.
4
Bedrock-only and text output only. Nova Premier is available exclusively through Amazon Bedrock, so there is no direct multi-cloud provider parity, and it outputs text only—no image, video, or audio generation despite accepting image and video input. Teams needing generative media output or a non-AWS deployment path must look elsewhere.

Best-Fit Workloads

Where this model earns its place

01

Long-document and long-video understanding

‍
The one-million-token window lets Premier ingest very long contracts, research collections, or up to roughly 90 minutes of video in a single prompt. Amazon documents strong performance on question answering over long documents. This suits legal review, policy analysis, and video-transcript Q&A where the entire source must be present. Structure inputs with long-form data first, since Amazon warns accuracy can decline slightly as context grows.

02

Whole-codebase analysis

‍
With a million-token context and documented strength in code understanding, Premier can reason over large codebases in one request rather than chunking. Amazon reports 42.4% on SWE-bench Verified—useful but behind leading Claude Sonnet versions—so it fits repository comprehension, cross-file summarization, and documentation over top-tier autonomous coding. For hard end-to-end code generation, pair or replace it with a stronger coding model.

03

Model distillation teacher

‍
Premier is explicitly designed as the teacher for Amazon Bedrock Model Distillation, producing cheaper, faster, lower-latency custom variants of Nova Pro, Lite, and Micro for specific verticals. This is a distinctive role: rather than serving Premier in production directly, teams use it offline to generate high-quality training signal, then deploy the distilled smaller model. It is the clearest case where paying for Premier's capability pays off downstream.

04

Multistep agentic and RAG workflows

‍
Premier supports tool calling and structured JSON-schema outputs via the Bedrock Converse API and is positioned for RAG, function calling, and multistep agentic execution across tools and data sources. The large context helps agents keep extensive reference data and history in-window. It fits enterprise agents that must reason over large grounding corpora—though latency and cost mean latency-critical agents may prefer a smaller Nova model.

Pricing in anytokens via AnyAPI
Input
15
₳
Output
75
₳
Cache write
—
₳
Cache read
—
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access Nova Premier via AnyAPI.ai

Access Nova Premier through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Nova Premier without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access Nova Premier and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test Nova Premier against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use Nova Premier from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use Nova Premier for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

Nova Premier 1.0 supports a one-million-token context window. Amazon describes this as roughly 1M text tokens, 500 images, or about 90 minutes of video in a single prompt, making it suited to large codebases, long documents, and long video. Amazon notes accuracy can decline slightly as the context grows.

Amazon's launch materials and independent analysis describe Nova Premier as a direct-response model, not an extended chain-of-thought reasoning model. It answers without a separate, controllable thinking budget. Note that the Amazon Bedrock model card lists reasoning as supported, which conflicts with launch positioning, so verify behavior for your specific use case.

Nova Premier accepts text, image, and video input, and produces text output only. It does not accept audio input and does not generate images, video, or audio. This makes it an understanding model for multimodal input rather than a generative-media model.

Both are multimodal Nova understanding models in Amazon Bedrock. Nova Premier offers a one-million-token context window and serves as the distillation teacher, while Nova Pro uses a 300K-token window and is tuned for faster, more cost-efficient production. Choose Premier for genuine long-context needs; choose Pro for most everyday workloads within 300K tokens.

Nova Premier is available in Amazon Bedrock through the Converse and Invoke APIs, with tool calling, structured outputs, streaming, and prompt caching. You can also access Nova Premier through AnyAPI.ai's unified API, calling it alongside other models for long-context and multimodal-input tasks.