Amazon Nova Pro

Amazon's balanced multimodal Bedrock model for text, image, and video at scale.

Context window

300K tokens

Input / 1M tokens

$0.80

Output / 1M tokens

$3.20

Provider

Amazon Web Services

$0.80 per 1M input tokens and $3.20 per 1M output tokens on Amazon Bedrock; cached input is about $0.20 per 1M. See the Amazon Bedrock pricing page for current rates. · Data verified 2026-07-02

Amazon Nova Pro is AWS's balanced multimodal foundation model, launched December 5, 2024 and available on Amazon Bedrock. It accepts text, image, and video input and produces text output, with a 300K-token context window and up to 5K output tokens. It targets a strong balance of accuracy, speed, and cost across a wide range of enterprise tasks, and supports Bedrock features like prompt caching, multiple service tiers, and broad regional availability. Knowledge cutoff is October 2024.

Capability index

Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.

Coding

Reasoning

Math

Multimodal

Long context

Speed

Cost efficiency

How to access it

Access via Amazon Bedrock (bedrock-runtime) using the Invoke or Converse API with model id 'amazon.nova-pro-v1:0' (or geo ids like us.amazon.nova-pro-v1:0). Authenticate with a Bedrock API key or AWS credentials and the boto3 SDK.

Get access →Documentation →

Strengths

✓Multimodal input across text, image, and video
✓Large 300K-token context window
✓Balanced accuracy/speed/cost tuned for enterprise workloads
✓Deep AWS/Bedrock integration (prompt caching, service tiers, wide region coverage)
✓Competitive pricing ($0.80/$3.20 per 1M) with cached-input discounts

Best for developers who...

AWS/Bedrock-native applicationsMultimodal text/image/video tasksCost-balanced general workloadsEnterprise deployments needing data residency

When to choose it (and when not to)

Reach for Amazon Nova Pro when...

→You are building on AWS and want native Bedrock integration
→You need multimodal understanding of images or video plus text
→You want a cost-balanced general model with a large context
→You need enterprise controls: data residency, service tiers, prompt caching

Look elsewhere if...

✕You need long generated outputs (max output is only 5K tokens)
✕You want a top-of-leaderboard frontier reasoning/coding model
✕You are not on AWS and don't want Bedrock as a dependency
✕You need audio input or image/video output

How to use it

›Use the Converse API for multi-turn and multimodal (image/video) prompts
›Keep outputs within the 5K-token cap; chunk long generations across calls
›Enable prompt caching for repeated system/context prefixes (text prompts, up to 20K cached tokens)
›Provide clear instructions and, for vision, high-quality media inputs

Quickstart

Python

import boto3

client = boto3.client("bedrock-runtime", region_name="us-east-1")

resp = client.converse(
    modelId="amazon.nova-pro-v1:0",
    messages=[{"role": "user", "content": [{"text": "Summarize the key features of Amazon Bedrock."}]}],
)
print(resp["output"]["message"]["content"][0]["text"])

Requires boto3 and AWS credentials (or a Bedrock API key via AWS_BEARER_TOKEN_BEDROCK). Model must be enabled in your Bedrock region.

API model id: amazon.nova-pro-v1:0

Benchmarks

Benchmark	Score	Notes
MMLU	~85%	Competitive generalist performance

Source: Amazon Nova Pro model card (AWS Bedrock docs)

Compare Amazon Nova Pro with

Amazon Nova Pro vs Gemini 2.5 Pro

Google DeepMind - 1,048,576 tokens (1M) input; up to 65K output ctx

Compare →

Amazon Nova Pro vs Claude Sonnet 4.6

Anthropic - 1M ctx

Compare →

Amazon Nova Pro vs GPT-4o

OpenAI - 128,000 tokens (16,384 max output) ctx

Compare →

All model comparisons →

Learn the concepts

RAG (Retrieval-Augmented Generation)Multimodal Context Window Embeddings

← All AI models