Amazon BedrockでAmazon Nova 1からAmazon Nova 2への移行

If you’re running Amazon Nova 1 models on Amazon Bedrock, you might be looking to expand your context window size, deepen reasoning capabilities, or integrate external tools for web search and code execution. Amazon Nova 2 models address these constraints while improving performance on reasoning, agentic AI, and tool use benchmarks.

Amazon Nova 2 Lite achieves higher scores across problem identification, solution completeness, and logical coherence benchmarks, while maintaining fast response times for high-volume workloads. If you’re using Nova 1 Lite for customer support automation, document processing, or agentic AI applications, you will likely see measurable gains in accuracy and throughput upon migrating to Nova 2 Lite. This post focuses on migration to Nova 2 Lite, which is generally available and ready for production use. Nova 2 expands the context window to one million tokens, which supports richer in-context learning and allows for processing longer documents in a single request. You also gain access to extended thinking, built-in web grounding, and a code interpreter. These features can enhance your existing applications with minimal code changes.

In this post, you will learn how to migrate from Nova 1 to Nova 2 on Amazon Bedrock. We cover model mapping, API changes, code examples using the Converse API, guidance on configuring new capabilities, and a summary of use cases. We conclude with a migration checklist to help you plan and execute your transition.

Migration paths

The following are recommended migration paths for the Nova 1 models. For Pro and Premier migrations, evaluate with extended thinking enabled to verify quality on your workloads.

• If you’re running Amazon Nova 1 Lite, the migration path is straightforward: Nova 2 Lite is a direct upgrade that maintains the same input modalities (text, image, and video) while adding extended thinking, built-in tools, and a context window that expands from 300K to 1M tokens.

• If you’re running Amazon Nova 1 Pro, we recommend upgrading to Nova 2 Lite. While this might seem like a tier change, Nova 2 Lite delivers improved reasoning capabilities at competitive price-performance. The extended thinking feature in Nova 2 Lite, combined with its 1M token context window, enables it to handle workloads that previously required Nova 1 Pro’s larger model size.

• If you’re running Amazon Nova Premier, consider migrating to Nova 2 Lite, especially for agentic and tool use workloads. As documented in the Amazon Nova 2 Technical Report, Nova 2 Lite surpasses Premier in multi-step problem-solving at 7x lower cost and up to 5x faster inference. Because Nova 2 Lite with extended thinking enabled is still cheaper and faster than Premier, we recommend evaluating across reasoning effort levels to verify that it meets your quality requirements.

Use cases and customer stories

Since Nova 2 Lite launched at re:Invent 2025, you can use the model for a variety of use cases. Nova 2 Lite excels at driving high throughput applications that require a combination of price, performance, and speed. Common use cases that customers have deployed Nova 2 Lite in production include:

Natural Language Processing (NLP) tasks

Nova 2 Lite is used for a wide variety of NLP tasks, from simple text predictors to logic-based recommendations. The Nova 2 models excel at summarization, classification, and search based use cases. Nova 2 Lite scores 80.9% on MMLU Pro and 70.8% on IF-Bench (with both higher than some comparable models). Nova 2 Lite is also designed for low latency, making it suitable for applications requiring rapid responses, as is typical with most NLP tasks.

Siemens global search runs on Nova 2 Lite, providing a 300% improvement in search speed and 70% cost reductions compared to their prior large language model (LLM) solution. The full case study can be found on the AWS case studies page.

Intelligent Document Processing (IDP)

Nova 2 models with reasoning enhance IDP by moving beyond basic text extraction to true semantic understanding of documents. They can process unstructured and semistructured content such as contracts, invoices, forms, and reports, and determine what information means in context, not only where it appears. Reasoning enables these models to link related information across sections, pages, and tables, handle variations in layout and language, and infer missing or implicit details, making document processing more robust and scalable. Reasoning adds value by (1) understanding context. For example, distinguishing a contract’s start date from an end date even if both appear as similar timestamps. (2) linking and inferring across pages. For example, tying invoice line items to totals/taxes and checking the math (even filling small gaps when text is unreadable). (3) normalizing variation—treating “amount payable,” “balance due,” and “total outstanding” as the same concept. (4) enabling smarter actions like validation, anomaly detection, and Q&A—flagging mismatched tax calculations or identifying auto-renewal clauses from related sections. Reasoning turns IDP into a smarter, more human-like system that understands and validates documents, enabling more accurate automation, less manual review, and faster adaptation to new ontology.

Multi-step agentic workflows

Nova 2 Lite shows considerable improvement over the Nova 1 models in tool calling capabilities, especially if multiple tools are part of the workflow. Overall, the Nova 2 family excels at completing complex, multi-step agentic workflows with sophisticated tool use and planning capabilities. Nova 2 Lite scores higher than comparable models in τ2-bench Telecom, at 76.0%. We recommend testing multi-step agentic use cases with reasoning on, starting with reasoning set to “Low” and then working upwards, if required.

Nova 2 Lite has empowered Trellix to continue to automate its security alert triage tasks and better empower security teams, in the areas of threat detection, analysis, and classification. With the release of the new model, Trellix saw more reliable tool calling capabilities with no failures, a 39% accuracy boost in threat classification, and 3.4x more detailed responses with technical analysis. Watch Trellix’s CTO explain the solution in this video.

AWS Transform is a multi-agent, multiuser system that uses dozens of tools to modernize complex infrastructure systems and code bases. AWS Transform breaks down each user request into a number of agent and tool calls that are routed to a mixture of models to balance cost, speed, and accuracy. Evaluations showed that Nova 2 Lite improves tool calling efficiency for code modernization in AWS Transform by up to 60%.

What’s changed

The Nova 2 family introduces capabilities that expand how you can build AI applications, particularly for agentic AI workloads.

Extended thinking with developer controls

Nova 2 Lite can reason through complex problems before responding. You control the depth of reasoning through the reasoningConfig parameter with low, medium, or high effort levels.

Native tool use and built-in web grounding and code interpreter

Nova 2 Lite is designed for tool use, including support for MCP servers and parallel tool chaining. It also includes built-in web grounding for real-time information with citations and a code interpreter for executing Python code directly.

Extended context for large-scale inputs

The context window expands from 300K to 1M tokens, and maximum output tokens increase from 10K to 65K. With these increases, you can process larger documents, code bases, and longer multi-turn workflows in a single request.

The following table summarizes the technical specifications across Nova 1 models and Nova 2 Lite:

Specification

Nova 1 Lite

Nova 1 Pro

Nova 2 Lite

Context Window

300K

300K

1M

Max Output Tokens

10K

10K

65K

Input Modalities

Text, image, video

Text, image, video

Text, image, video

Extended Thinking

No

No

Yes

Built-in Tools

No

No

Yes

Customization

Yes (incl. Nova Forge)

Yes (incl. Nova Forge)

Yes (incl. Nova Forge)

Performance and pricing

Independent benchmarks from Artificial Analysis show that Nova 2 Lite scores above average on intelligence among comparable models and is notably fast, making it well-suited for high-volume workloads where throughput matters. On agentic-specific benchmarks, Nova 2 Lite performs well on Tool Use (Tau2 Telecom). Strong instruction following is critical for agentic workloads where the model must reliably execute multi-step plans. For detailed benchmark methodology and additional results, see the Amazon Nova 2 Technical Report.

Pricing for the Nova 2 family reflects the expanded capabilities while remaining competitive within their respective tiers.

Model

Input (per 1M tokens)

Output (per 1M tokens)

Nova 1 Lite

$0.06

$0.24

Nova 1 Pro

$0.80

$3.20

Nova 2 Lite

$0.30

$2.50

For Nova 2 Lite, pricing is competitive against similar models in its class. For the full announcement and additional details, see the Amazon Nova 2 launch post.

API changes and code updates

Update your model ID to reference Nova 2 Lite, which is available through Global Cross-Region Inference System (CRIS), US CRIS, EU CRIS, and JP CRIS endpoints. When accessing models from a US region, use the us. prefix (for example, us.amazon.nova-2-lite-v1:0); when accessing from outside the US, use the global. prefix (for example, global.amazon.nova-2-lite-v1:0). The underlying API structure remains consistent, so existing integrations will work with minimal changes beyond the model ID and any new features that you choose to enable.

Breaking change: When maxReasoningEffort is set to high, you cannot use maxTokens, temperature, topP, or topK. Requests that include these parameters with high effort reasoning will return an error. To resolve this, remove the inferenceConfig block entirely from your request.

New features and unchanged behaviors

The following table summarizes new features and unchanged behaviors between Nova 1 and Nova 2:

Category

Change

Details

New

Extended thinking

Enable reasoning using reasoningConfig in additionalModelRequestFields. Supports three effort levels: low, medium, and high.

New

Reasoning content in responses

Responses include a reasoningContent field when reasoning is enabled. Content displays as [REDACTED] but reasoning tokens are billed.

New

Built-in tools

Web grounding and code interpreter are available without external integrations.

New

Increased output capacity

Maximum output tokens increased from 10K to 65,536.

New

Expanded context window

Nova 2 Lite supports 1M tokens (up from 300K in Nova 1 Lite).

Unchanged

topK parameter location

topK is still passed through additionalModelRequestFields, not inferenceConfig.

Unchanged

Document support

Document input is still Converse API only (not supported in Invoke API).

Unchanged

Timeout configuration

Boto3’s default read timeout is 60 seconds, but extended thinking requests can take up to 60 minutes. Configure read_timeout=3600 in your SDK client.

Configuring new capabilities

This section covers the basic migration steps, extended thinking configuration, and how to use built-in tools like web grounding and code interpreter.

Basic migration – update the model ID

For applications that do not require extended thinking, the migration is straightforward. Update the model ID and your existing code will work:

import boto3

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

# Nova 1
# modelId='us.amazon.nova-lite-v1:0'

# Nova 2
response = bedrock.converse(
    modelId='us.amazon.nova-2-lite-v1:0',
    system=[{'text': 'You are a helpful assistant'}],
    messages=[
        {
            'role': 'user',
            'content': [{'text': 'Explain cloud computing in simple terms.'}]
        }
    ],
    inferenceConfig={
        'maxTokens': 1024,
        'temperature': 0.7,
        'topP': 0.9
    }
)

print(response['output']['message']['content'][0]['text'])

Extended thinking

Extended thinking allows Nova 2 model to reason through complex problems before generating responses. This capability is disabled by default to optimize for speed and cost on simple queries. When your workload requires multi-step reasoning, you can enable it through the reasoningConfig parameter.

Configuring reasoning effort levels

Nova 2 Lite provides three effort levels that control how much reasoning the model performs:

Low effort (maxReasoningEffort: "low") is recommended for tasks with added complexity that benefit from structu