Amazon Nova API: Access, Rate Limits & SDKs 2026
Nova 2 is not a standalone API but foundation models on Amazon Bedrock. You use AWS credentials and call the unified Converse API, with no servers to run.
Amazon Nova API verdict
Nova 2 is not a standalone API but a set of foundation models on Amazon Bedrock.
You authenticate with AWS credentials, whether IAM users, roles or Identity Center. Then you attach Bedrock permissions like AmazonBedrockFullAccess plus a marketplace Subscribe policy, and call the unified Converse API with a regional model id such as us.amazon.nova-2-pro-v1:0.
If you are on AWS, Nova 2's API story is a strength. One Converse operation runs across boto3 and every AWS SDK, with no infrastructure to run. You also get IAM and CloudTrail for governance, built-in web grounding and a code interpreter, and easy RAG over data already in AWS. Wire it up with AWS credentials, attach AmazonBedrockFullAccess plus a marketplace Subscribe policy, and call us.amazon.nova-2-pro-v1:0. Add reasoningConfig for extended thinking, or switch to Nova 2 Lite for cheap high-volume work and Sonic for voice. If you are not on AWS, the friction is real. There is no simple bearer-key endpoint, access is IAM- and subscription-gated, and limits are Bedrock service quotas, so a model with a direct API is lower-friction.
- Nova 2 is AWS-only. It runs on Amazon Bedrock and needs an AWS account, IAM credentials and Bedrock permissions, not a simple bearer key.
- Access is gated twice: AmazonBedrockFullAccess or scoped permissions, plus a marketplace Subscribe policy to enable the models.
- Model ids are regional inference profiles, like us.amazon.nova-2-pro-v1:0, so use the right region prefix or the call fails.
- Rate and throughput limits are Amazon Bedrock service quotas, per account and region, raised through AWS Service Quotas rather than a vendor dashboard.
- Access
- Amazon Bedrock (Converse API)
- Auth
- AWS IAM credentials
- Model id
- us.amazon.nova-2-pro-v1:0
- Built-in tools
- Web grounding + code interpreter
- SDKs
- boto3 + AWS SDKs
Build with the Amazon Nova API
import boto3
client = boto3.client("bedrock-runtime", region_name="us-east-1")
response = client.converse(
modelId="us.amazon.nova-2-pro-v1:0",
messages=[{"role": "user", "content": [{"text": "Hello, Nova"}]}],
)
print(response["output"]["message"]["content"][0]["text"]){
"output": { "message": { "role": "assistant",
"content": [{ "text": "Hello! How can I help?" }] } },
"stopReason": "end_turn",
"usage": { "inputTokens": 5, "outputTokens": 8 }
}- Cost levers: Nova 2 Lite for cheap high-volume work; concise prompts (output is the expensive token); built-in tools instead of external glue code.
- Handle: ThrottlingException (Bedrock service quota exceeded), AccessDeniedException (missing Bedrock permission / model not subscribed).
Real Amazon Nova code from the official docs. The raw cURL hello-world is the code block above; full limits and endpoints are in the tables on this page.
Your first Amazon Nova API request
aws bedrock-runtime converse \
--model-id us.amazon.nova-2-pro-v1:0 \
--messages '[{"role":"user","content":[{"text":"Write a short poem"}]}]'How to authenticate with the Amazon Nova API
| Aspect | Detail | Notes |
|---|---|---|
| Auth | AWS credentials (IAM users, roles, Identity Center) | boto3 / AWS CLI / SDKs sign requests |
| Permissions | AmazonBedrockFullAccess managed policy | Plus marketplace Subscribe to enable models |
| Service | Amazon Bedrock (bedrock-runtime) | Foundation models with no infrastructure to manage |
| Model id | us.amazon.nova-2-pro-v1:0 | Regional prefix; Lite is us.amazon.nova-2-lite-v1:0 |
| Primary call | Converse API (bedrock-runtime converse) | Unified messages interface across models |
| Model family | Nova 2 Pro, Lite, Sonic, Multimodal Embeddings | Pick by modality and cost |
| Deployment | Bedrock managed, or fine-tune / customize | Customize on SageMaker AI / Nova Forge |
Core capabilities and tools
| Capability | How | Purpose |
|---|---|---|
| Converse | bedrock-runtime converse | Generate from text or multimodal input (unified messages) |
| Reasoning toggle | reasoningConfig: {type, maxReasoningEffort} | Enable extended step-by-step thinking |
| Multimodal input | content blocks: text, image, video, documents | Understand media alongside text in one request |
| Web grounding | Built-in tool | Access real-time information without external integrations |
| Code interpreter | Built-in tool | Execute Python code without external integrations |
| RAG | Incorporate your data sources | More accurate, contextual answers from your data |
| Streaming | ConverseStream | Stream the response as it is generated |
| Customization | Fine-tuning / RL on SageMaker AI, Nova Forge | Adapt models to your use case |
Quotas, regions and model rules
| Limit | Detail | Notes |
|---|---|---|
| Max output | Up to 65,536 tokens per response | Nova 2 family generation limit |
| Context | 256k (Pro entry); up to 1M for the family | Whole codebases / long documents |
| Regional model id | us. prefix (e.g. us.amazon.nova-2-pro-v1:0) | Cross-region inference profile |
| Model subscription | Marketplace Subscribe permission required | Inline policy for aws-marketplace:Subscribe |
| Permissions | AmazonBedrockFullAccess (or scoped) | Attach to the user or role |
| Quotas | Amazon Bedrock service quotas apply | Per account / region; raise via Service Quotas |
| Credentials | IAM long-term or temporary (roles / SSO) | Sign requests via AWS CLI / SDKs |
SDKs, ecosystem and developer fit
| Aspect | Detail | Notes |
|---|---|---|
| SDKs | boto3 (Python), AWS SDKs, AWS CLI | Converse / ConverseStream operations |
| Built-in tools | Web grounding + code interpreter | No external integrations needed |
| Multimodal build | Document assistant, video analysis, voice agent | Nova 2 Lite (video), Sonic (speech) |
| Customization | Fine-tuning + reinforcement learning | On SageMaker AI; Nova Forge |
| Workflow | App -> Bedrock -> Nova -> response | Optionally RAG + tools + reasoning |
| Embeddings | Nova Multimodal Embeddings | Semantic search, recommendations, similarity |
| API pricing (Nova 2 Pro) | $1.25/1M in, $10.00/1M out | Independent cross-reference; Bedrock token pricing |
Amazon Nova API FAQ
How do I authenticate with the Nova 2 API?
Nova 2 runs on Amazon Bedrock, so you use AWS credentials, not a bearer key. Set up programmatic access, whether IAM user long-term keys, an IAM role, or IAM Identity Center temporary credentials. Attach the AmazonBedrockFullAccess managed policy plus a marketplace Subscribe inline policy, then call the Converse API with boto3 or the AWS CLI against a model id like us.amazon.nova-2-pro-v1:0.
What is the Converse API and what can it do?
Converse is Bedrock's unified messages interface. For Nova 2 it handles text and multimodal input across image, video, documents and speech, plus streaming through ConverseStream. It also supports a reasoningConfig toggle for extended thinking, tool use, and Amazon's built-in web grounding and code interpreter. RAG over your own data runs through the same operation.
Which Nova 2 models are available and how do I pick?
Nova 2 Pro is the multimodal flagship for complex tasks. Nova 2 Lite is optimized for high-volume, cost-sensitive work and video. Nova 2 Sonic handles speech-to-speech voice applications, and Nova Multimodal Embeddings powers semantic search and recommendations. All are reached through Bedrock with the same Converse interface and regional model ids.
What are the Nova 2 API limits?
The model generates up to 65,536 tokens per response and supports 256k tokens of context on the Pro entry, up to 1M across the family. Access limits are AWS-shaped. Amazon Bedrock service quotas apply per account and region, raised through AWS Service Quotas, and you need both IAM permissions and a marketplace model subscription before you can invoke a model.
Can I customize or fine-tune Nova 2?
Yes. Beyond calling the base models through Bedrock, you can customize Nova 2 through fine-tuning and reinforcement learning on Amazon SageMaker AI, and through Nova Forge. That adapts a model to your domain and tasks. You can also enhance responses without training, by enabling RAG over your data and the built-in web grounding and code-interpreter tools.
Sources & verification
| Source | What was checked | Last checked |
|---|---|---|
| Amazon Developer docs | Official product page | July 10, 2026 |
| Amazon Developer docs | Nova2 Userguide Getting Started API | July 10, 2026 |
| Artificial Analysis | Independent reference | July 10, 2026 |
Every fact on this Amazon Nova page is tied to a named source and a verification date. Freshness-sensitive figures trace to the sources above; verify against the vendor before relying on them.
Explore Amazon Nova
Every page on Amazon Nova in one place, you are on api.