TheRouter.aiTheRouter.ai
Back to Models

Amazon Nova Premier

amazonamazon/nova-premier

Amazon's most capable multimodal model for complex reasoning tasks. Best teacher for distilling custom models. Supports text, images, and videos.

AWS What's New announcement ↗Bedrock model card ↗Nova product page ↗Amazon Nova Premier technical report (amazon.science) ↗

Amazon Nova Premier was Amazon's most capable multimodal foundation model at launch — built for complex reasoning, long-context processing, and agentic workflows. Released into general availability on April 30, 2025 (Bedrock model launch date: October 31, 2025), it accepts text, images, and video inputs in a one-million-token context window. It carries a Legacy lifecycle tag and an EOL date of September 14, 2026, meaning AWS has already moved to the Nova 2 generation for new production builds.

The defining feature AWS highlighted at launch was model distillation: Nova Premier is the designated teacher model for creating smaller, faster versions of Nova Pro, Lite, and Micro via Amazon Bedrock Model Distillation. In their own example, distilling Nova Pro for tool selection and API calling produced a model with 20% higher API invocation accuracy that consistently matched the teacher at Nova Pro's speed and cost. For TheRouter operators, this is primarily relevant if you run Bedrock-native distillation pipelines; for standard API routing, the Legacy status makes Nova 2 Lite the current Amazon offering to reach for.

Best for
  • • Amazon Bedrock Model Distillation — creating custom, cost-optimized Nova Pro/Lite/Micro variants for narrow production workflows
  • • Legacy Bedrock workloads already running on Nova Premier that cannot be migrated before the September 14, 2026 EOL
  • • Long-document and large-codebase analysis requiring a 1M context with multimodal (text + image + video) input
Reach for something else if
  • • New agentic or reasoning workloads — Nova Premier reaches EOL September 14, 2026; build on amazon/nova-2-lite instead, which offers a comparable 1M context with extended-thinking budgets and an active lifecycle
  • • Workloads needing structured JSON outputs — Nova Premier does not support Bedrock Structured Outputs; use a model that does if strict schema adherence is required
  • • Audio-in or audio-out pipelines — Nova Premier accepts text, image, and video input only; use a dedicated speech model for audio

How TheRouter serves this differently from the vendor

As the vendor operates it

AWS serves Nova Premier through Bedrock bedrock-runtime with Converse, InvokeModel, response streaming, guardrails, prompt caching (1K min / 4 checkpoints / 5-min TTL), and client-side tool calling. Structured Outputs are not supported. Global cross-region inference is not supported — only Geo inference (us.amazon.nova-premier-v1:0) and in-region (amazon.nova-premier-v1:0). Lifecycle is Legacy with EOL September 14, 2026.

On TheRouter

TheRouter exposes Nova Premier as OpenAI-compatible /v1/chat/completions under amazon/nova-premier. Operational routing, pricing, context, modality, and supported-parameter fields are taken from the live TheRouter catalog. Snippets on this page are unverified until an operator-funded SNIPPET_VERIFY_APPROVED run completes; completeness is capped accordingly.

docs.aws.amazon.com/bedrock ↗
Context Length
1M
Max Output
5K
Input Priceper 1M tokens
$2.70/ 1M tokens
Output Priceper 1M tokens
$13.50/ 1M tokens

Modalities

textimage→text

Capabilities

Vision

Pricing Breakdown

TypeRate
Input$2.70 / 1M tokens
Output$13.50 / 1M tokens

Supported Parameters

temperaturemax_tokenstop_ptoolstool_choiceresponse_formatstop

Specifications

General availability dateApril 30, 2025aws.amazon.com ↗verified
Bedrock model launch dateOctober 31, 2025docs.aws.amazon.com ↗verified
Model lifecycleLegacydocs.aws.amazon.com ↗verified
EOL dateSeptember 14, 2026docs.aws.amazon.com ↗verified
Knowledge cutoffOctober 2024docs.aws.amazon.com ↗verified
Max output tokens25,000 tokensdocs.aws.amazon.com ↗verified
Reasoning (extended thinking)Supporteddocs.aws.amazon.com ↗verified
Prompt cachingSupported on system and messages fields; 1K minimum per checkpoint, up to 4 checkpoints, 5-minute TTLdocs.aws.amazon.com ↗verified
Structured outputsNot supporteddocs.aws.amazon.com ↗verified
Bedrock model IDamazon.nova-premier-v1:0 (also Geo ID: us.amazon.nova-premier-v1:0; Global inference: not supported)docs.aws.amazon.com ↗verified
LicenseAWS Service Terms / Bedrock Third-Party Model EULA (proprietary, API-only via Bedrock)aws.amazon.com ↗verified

Benchmarks

BenchmarkDistributionScoreSource
MMLU
Massive Multitask Language Understanding — undergraduate-level knowledge across 57 subjects. AWS published this score in the April 30, 2025 GA announcement.
87.4%%aws.amazon.com ↗
Math500
Math500 benchmark for mathematical problem solving. AWS published this score in the April 30, 2025 GA announcement.
82.0%%aws.amazon.com ↗
CharXiv
CharXiv benchmark for chart understanding. AWS published this score in the April 30, 2025 GA announcement.
84.6%%aws.amazon.com ↗
SimpleQA (with RAG)
Retrieval-Augmented Generation accuracy on SimpleQA. AWS published this score in the April 30, 2025 GA announcement.
86.3%%aws.amazon.com ↗
Berkeley Function Calling Leaderboard (BFCL)
Function calling accuracy on BFCL. AWS published this score in the April 30, 2025 GA announcement.
63.7%%aws.amazon.com ↗
SWE-bench Verified
Software engineering task resolution on SWE-bench Verified. AWS published this score in the April 30, 2025 GA announcement.
42.4%%aws.amazon.com ↗

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

cURL
curl https://api.therouter.ai/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -d '{
    "model": "amazon/nova-premier",
    "messages": [
      {"role": "user", "content": "Summarize the key points from this input."}
    ]
  }'

API guide

Full API reference →

Chat completion

Call Nova Premier through TheRouter's OpenAI-compatible surface. TheRouter handles the Bedrock InvokeModel / Converse translation — your client stays OpenAI-shaped. Note the Legacy lifecycle: plan a migration to amazon/nova-2-lite before September 14, 2026.

cURL
curl https://api.therouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "amazon/nova-premier",
    "messages": [{"role": "user", "content": "Analyse this 90-minute video transcript and identify key decisions made."}]
  }'

More from amazon

amazon/nova-2-liteActive replacement — 1M context, extended-thinking budgets, web grounding built in. Migrate here before Nova Premier EOL.amazon/nova-proMid-tier Nova — 300K context, broad multimodal support, active lifecycle.amazon/nova-microCheapest, text-only Nova — for high-volume classification and short-form completion.

Similar models

Cross-provider sibling models
anthropic/claude-opus-4.7
Frontier reasoning and coding model — strong alternative for complex agentic tasks with an active lifecycle.
google/gemini-2.5-pro
Google's top reasoning model — comparable long-context and multimodal capabilities at a different price point.

News & changes

2025-04-30

Amazon Nova Premier reaches general availability — 1M context, distillation teacher, SWE-bench 42.4%

AWS announced general availability of Nova Premier on April 30, 2025, positioning it as the highest-capability model in the Nova family at launch. Key operator-relevant hooks: 1M context for large codebase and document ingestion; MMLU 87.4%, SWE-bench 42.4%; and the model distillation role — using Nova Premier as teacher to create custom, cost-optimised variants of Nova Pro, Lite, and Micro. AWS reported that a distilled Nova Pro for API calling delivered 20% higher accuracy at Nova Pro speed and cost. Available via Bedrock cross-region inference in US East and US West at launch.

re-authored by TheRouteraws.amazon.com/whats-new ↗

Frequently asked

Is Nova Premier still available?

Yes, but it carries a Legacy lifecycle status on AWS Bedrock with an EOL date of September 14, 2026. TheRouter continues to route requests to it until that date, but new workloads should be built on amazon/nova-2-lite or another active model. Existing integrations should plan a migration before the EOL.

docs.aws.amazon.com ↗
What is model distillation and how does Nova Premier support it?

Amazon Bedrock Model Distillation lets you use a powerful model (the teacher) to generate training data that trains a smaller, cheaper model (the student) for a specific task. Nova Premier is the designated teacher for distilling Nova Pro, Lite, and Micro. AWS showed that distilling Nova Pro for API calling produced a model with 20% higher accuracy at Nova Pro's speed and cost — a compelling argument if you need heavy reasoning quality at lighter-tier latency and pricing.

aws.amazon.com ↗
What should I migrate to before the September 14, 2026 EOL?

For agentic and long-context workloads, amazon/nova-2-lite is the natural successor: it has a 1M context window, built-in extended-thinking budgets (low/medium/high), web grounding and a code interpreter as runtime tools, prompt caching, and an active lifecycle. If you need a higher-accuracy tier, wait for Nova 2 Pro's GA announcement or route to anthropic/claude-opus-4.7.

docs.aws.amazon.com ↗
Does Nova Premier support structured JSON output?

No. The Bedrock model card explicitly lists Structured Outputs as not supported for Nova Premier. You can prompt-engineer JSON-shaped output with a system prompt and response_format: json_object, but schema enforcement at the API level is not available. Consider models that natively support structured outputs if strict schema adherence is required.

docs.aws.amazon.com ↗
What are the benchmark scores for Nova Premier?

AWS published the following at GA on April 30, 2025: MMLU 87.4% (undergraduate knowledge), Math500 82.0% (mathematics), CharXiv 84.6% (chart understanding), SimpleQA with RAG 86.3% (retrieval accuracy), BFCL 63.7% (function calling), SWE-bench Verified 42.4% (software engineering). These scores are from the AWS GA announcement and have not been independently re-verified on TheRouter's infrastructure.

aws.amazon.com ↗
Fact ledger — every claim on this page traces here
sourceURLretrieved
General availability dateaws.amazon.com ↗2026-08-18verified
Bedrock model launch datedocs.aws.amazon.com ↗2026-08-18verified
Model lifecycledocs.aws.amazon.com ↗2026-08-18verified
EOL datedocs.aws.amazon.com ↗2026-08-18verified
Knowledge cutoffdocs.aws.amazon.com ↗2026-08-18verified
Max output tokensdocs.aws.amazon.com ↗2026-08-18verified
Reasoning (extended thinking)docs.aws.amazon.com ↗2026-08-18verified
Prompt cachingdocs.aws.amazon.com ↗2026-08-18verified
Structured outputsdocs.aws.amazon.com ↗2026-08-18verified
Bedrock model IDdocs.aws.amazon.com ↗2026-08-18verified
Licenseaws.amazon.com ↗2026-08-18verified
MMLUaws.amazon.com ↗2026-08-18verified
Math500aws.amazon.com ↗2026-08-18verified
CharXivaws.amazon.com ↗2026-08-18verified
SimpleQA (with RAG)aws.amazon.com ↗2026-08-18verified
Berkeley Function Calling Leaderboard (BFCL)aws.amazon.com ↗2026-08-18verified
SWE-bench Verifiedaws.amazon.com ↗2026-08-18verified
Amazon Nova Premier reaches general availability — 1M context, distillation teacher, SWE-bench 42.4%aws.amazon.com/whats-new ↗2026-08-18verified
Is Nova Premier still available?docs.aws.amazon.com ↗2026-08-18to verify
What is model distillation and how does Nova Premier support it?aws.amazon.com ↗2026-08-18to verify
What should I migrate to before the September 14, 2026 EOL?docs.aws.amazon.com ↗2026-08-18to verify
Does Nova Premier support structured JSON output?docs.aws.amazon.com ↗2026-08-18to verify
What are the benchmark scores for Nova Premier?aws.amazon.com ↗2026-08-18to verify
Help & contact