In the drive to remain competitive, businesses today are turning to AI to help them minimize cost and maximize efficiency. It’s incumbent on them to find the most suitable AI model—the one that will help them achieve more while spending less. For many businesses, the migration from OpenAI’s model family to Amazon Nova represents not only a shift in model but a strategic move toward scalability, efficiency, and broader multimodal capabilities.
In this blog, we discuss how to optimize prompting in Amazon Nova for the best price-performance.
Why migrate from OpenAI to Amazon Nova?
OpenAI’s models remain powerful, but their operational costs can be prohibitive when scaled. Consider these figures from Artificial Analysis:
| Model | Input Token Cost (per Million Tokens) | Output Token Cost (per Million Tokens) | Context Window | Output Speed (Tokens per Second) | Latency (Seconds per first token) |
| GPT-4o | ~$2.50 | ~$10.00 | Up to 128K tokens | ~63 | ~0.49 |
| GPT-4o Mini | ~$0.15 | ~$0.60 | Up to 128K tokens | ~90 | ~0.43 |
| Nova Micro | ~$0.035 | ~$0.14 | Up to 128K tokens | ~195 | ~0.29 |
| Nova Lite | ~$0.06 | ~$0.24 | Up to 300K tokens | ~146 | ~0.29 |
| Nova Pro | ~$0.80 | ~$3.20 | Up to 300K tokens | ~90 | ~0.34 |
For high-volume applications—like global customer support or large-scale document analysis—these cost differences are disruptive. Not only does Amazon Nova Pro offer over three times the cost-efficiency, its longer context window also enables it to handle more extensive and complex inputs.
Breaking down the Amazon Nova suite
Amazon Nova isn’t a single model—it’s a suite designed for various needs:
- Amazon Nova Pro – A robust multimodal model that can process text, images, and video. It excels at tasks such as document analysis and deep data visualization. Benchmark comparisons show Amazon Nova Pro matching or even surpassing GPT-4o on complex reasoning tasks, according to section 2.1.1 of the Nova technical report and model card.
- Amazon Nova Lite – Offers a balanced mix of multimodal processing and speed. Amazon Nova Lite is ideal for applications such as document summarization, translation, and even basic visual search, delivering quality outputs at lower latency and cost compared to GPT-4o Mini. You can find these benchmark results in section 2.1.2 of the Nova Technical Report and Model Card.
- Amazon Nova Micro – A text-only model engineered for ultra-low latency. With output speed of up to 195 tokens per second, Amazon Nova Micro is perfect for real-time applications such as chat-based assistants and automated FAQs. Its token costs are dramatically lower than those of GPT-4o Mini—roughly 4.3 times cheaper on a per-token basis.
The lower per-token costs and higher output per second of Amazon Nova give you the flexibility to simplify prompts for real-time applications so you can balance quality, speed, and cost for your use case.
Understanding the foundations
To make the best decision about which model family fits your needs, it’s important to understand the differences in prompt engineering best practices in both OpenAI and Amazon Nova. Each model family has its own set of strengths, but there are some things that apply to both families. Across both model families, quality accuracy is achieved through clarity of instructions, structured prompts, and iterative refinement. Whether you’re using strong output directives or clearly defined use cases, the goal is to reduce ambiguity and improve response quality.
The OpenAI approach
OpenAI uses a layered messaging system for prompt engineering, where system, developer, and user prompts work in harmony to control tone, safety, and output format. Their approach emphasizes:
- Hierarchical message roles – Setting the model’s role and behavior using system messages makes sure that the overarching safety and style guidelines (set in system prompts) are preserved
- Instruction placement and delimiters – Directives are placed at the beginning, with clear separation between context, examples, and queries
- Selective chain-of-thought – Detailed, step-by-step reasoning is used when it benefits complex tasks
- Formatting and structure – Using strong directives such as DO, MUST, and DO NOT to provide consistent outputs (for example, in JSON)
The Amazon Nova approach
- Define the prompt use case
- Task – What exactly the model should do
- Role – Which role the model should assume
- Response style – The structure or tone of the output
- Instructions – Guidelines the model must follow
- Chain-of-thought
- Explicit state response – Provide clear and strong instructions to limit the model’s response
- Structural thinking – Thinking step-by-step encourages structural thinking
- Formatting and structure
- Use delimiters to sectioned your prompts, for example, ##Task##, ##Context##, or ##Example##
- Specify the output format, for example, JSON, YAML, or Markdown
- Use strong instructions and caps, such as DO, DO NOT, or MUST
- Prefill the responses to guide the model, for example, start with “{“ or “json…”
Evolving your prompt strategy: Migrating from OpenAI to Amazon Nova
Transitioning to Amazon Nova isn’t merely a change in API endpoints—it requires retooling your prompt engineering to align with the strengths of Amazon Nova. You need to reframe your use case definition. Begin by breaking down your current GPT-4o or GPT-4o Mini prompt into its core elements of task, role, response style, and instructions and success criteria. Make sure to structure these elements clearly to provide a blueprint for the model.
To understand how to migrate an existing OpenAI prompt to work optimally for Amazon Nova Pro, consider the following example using the meeting notes summarizer. Here is the GPT-4o system prompt:
The user prompt is the meeting notes that need to be summarized:
GPT produces this helpful response:
To meet or exceed the quality of the response from GPT-4o, here is what an Amazon Nova Pro prompt might look like. The prompt uses the same best practices discussed in this post, starting with the system prompt. We used a temperature of .2 and a topP of .9 here:
Here’s the user prompt, using prefilled responses:
The following example shows that the Amazon Nova response meets and exceeds the accuracy of the OpenAI example, formats the output in Markdown, and has found clear owners for each action item:
A few updates to the prompt can achieve comparable or better results from Amazon Nova Pro while enjoying a much less expensive cost of inference.
Employ Amazon Nova extended context
Amazon Nova Lite and Amazon Nova Pro can support up to 300,000 input tokens, which means that you can include more context in your prompt if needed. Expand your background data and detailed instructions accordingly—if your original OpenAI prompt was optimized for 128,000 tokens, adjust it to use the Amazon Nova extended window.
Tailor output constraints
If your GPT prompt required strict formatting (for example, “Respond in JSON only”), make sure that your Amazon Nova prompt includes these directives. Additionally, if your task involves multimodal inputs, specify when to include images or video references.
Function calling
The rise of generative AI agents has made function calling, or tool calling, one of the most important abilities of a given large language model (LLM). A model’s ability to correctly pick the right tool for the job, in a low-latency manner, is often the difference between success and failure of an agentic system.
Both OpenAI and Amazon Nova models share similarities in function calling, in particular their support for structured API calls. Both model families support tool selection through defined tool schemas, which we discuss later in this post. They also both provide a mechanism to decide when to invoke these tools or not.
OpenAI’s function calling uses flexible JSON schemas to define and structure API interactions. The models support a wide range of schema configurations, which give developers the ability to quickly implement external function calls through straightforward JSON definitions tied to their API endpoints.
Here is an example of a function:
tools = [{
"type": "function",
"function": {
"name": "get_weather",
"description": "Get current temperature for a given location.",
"parameters": {
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "City and country e.g. Montevideo, Uruguay"
}
},
"required": [
"location"
],
"additionalProperties": False
},
"strict": True
}
}]
completion = client.chat.completions.create(
model="gpt-4o",
messages=[{"role": "user", "content": "What is the weather like in Punta del Este today?"}],
tools=toolsSimilar to OpenAI’s approach, Amazon Nova can call tools when passed a configuration schema as shown in the following code example. Amazon Nova has made heavy use of Greedy Decoding when calling tools, and it’s advised to set temperature, topP, and topK to 1. This makes sure that the model has the highest accuracy in tool selection. These Greedy Decoding parameters and other great examples of tool use are covered in great detail in Tool use (function calling) with Amazon Nova.
The following is an example of function calling without using additionalModelRequestFields:
tool_config = {
"tools": [{
"toolSpec": {
"name": "get_recipe",
"description": "Structured recipe generation system",
"inputSchema": {
"json": {
"type": "object",
"properties": {
"recipe": {
"type": "object",
"properties": {
"name": {"type": "string"},
"ingredients": {
"type": "array",
"items": {
"type": "object",
"properties": {
"item": {"type": "string"},
"amount": {"type": "number"},
"unit": {"type": "string"}
}
}
},
"instructions": {
"type": "array",
"items": {"type": "string"}
}
},
"required": ["name", "ingredients", "instructions"]
}
}
}
}
}
}]
}
# Base configuration without topK=1
input_text = "I need a recipe for chocolate lava cake"
messages = [{
"role": "user",
"content": [{"text": input_text}]
}]
# Inference parameters
inf_params = {"topP": 1, "temperature": 1}
response = client.converse(
modelId="us.amazon.nova-lite-v1:0",
messages=messages,
toolConfig=tool_config,
inferenceConfig=inf_params
)
# Typically produces less structured or incomplete output
The following example shows how function calling accuracy can be improved by using
additionalModelRequestFields:
# Enhanced configuration with topK=1
response = client.converse(
modelId="us.amazon.nova-lite-v1:0",
messages=messages,
toolConfig=tool_config,
inferenceConfig=inf_params,
additionalModelRequestFields={"inferenceConfig": {"topK": 1}}
)
# Produces more accurate and structured function call
To maximize Amazon Nova function calling potential and improve accuracy, always use additionalModelRequestFields with topk=1. This forces the model to select the single most probable token and prevents random token selection. This increases deterministic output generation and improves function call precision by about 30–40%.
The following code examples further explain how to conduct tool calling successfully. The first scenario shows recipe generation without an explicit tool. The example doesn’t use topK, which typically results in responses that are less structured:
input_text = """
I'm looking for a decadent chocolate dessert that's quick to prepare.
Something that looks fancy but isn't complicated to make.
"""
messages = [{
"role": "user",
"content": [{"text": input_text}]
}]
response = client.converse(
modelId="us.amazon.nova-lite-v1:0",
messages=messages,
inferenceConfig={"topP": 1, "temperature": 1}
)
# Generates a conversational recipe description
# Less structured, more narrative-driven response
In this example, the scenario shows recipe generation with a structured tool. We add topK set to 1, which produces a more structured output:
response = client.converse(
modelId="us.amazon.nova-lite-v1:0",
messages=messages,
toolConfig=tool_config,
inferenceConfig={"topP": 1, "temperature": 1},
additionalModelRequestFields={"inferenceConfig": {"topK": 1}}
)
# Generates a highly structured, JSON-compliant recipe
# Includes precise ingredient measurements
# Provides step-by-step instructionsOverall, OpenAI offers more flexible, broader schema support. Amazon Nova provides more precise, controlled output generation and is the best choice when working with high-stakes, structured data scenarios, as demonstrated in Amazon Nova’s performance on the IFEval benchmark discussed in section 2.1.1 of the technical report and model card. We recommend using Amazon Nova for applications requiring predictable, structured responses because its function calling methodology provides superior control and accuracy.
Conclusion
The evolution from OpenAI’s models to Amazon Nova represents a significant shift in using AI. It shows a transition toward models that deliver similar or superior performance at a fraction of the cost, with expanded capabilities in multimodal processing and extended context handling.
Whether you’re using the robust, enterprise-ready Amazon Nova Pro, the agile and economical Amazon Nova Lite, or the versatile Amazon Nova Micro, the benefits are clear:
- Cost savings – With token costs up to four times lower, businesses can scale applications more economically
- Enhanced response performance – Faster response times (up to 190 tokens per second) make real-time applications more viable
- Expanded capabilities – A larger context window and multimodal support unlock new applications, from detailed document analysis to integrated visual content
By evolving your prompt strategy—redefining use cases, exploiting the extended context, and iteratively refining instructions—you can smoothly migrate your existing workflows from OpenAI’s o4 and o4-mini models to the innovative world of Amazon Nova.
About the Authors
Claudio Mazzoni is a Sr Specialist Solutions Architect on the Amazon Bedrock GTM team. Claudio exceeds at guiding costumers through their Gen AI journey. Outside of work, Claudio enjoys spending time with family, working in his garden, and cooking Uruguayan food.
Pat Reilly is a Sr. Specialist Solutions Architect on the Amazon Bedrock Go-to-Market team. Pat has spent the last 15 years in analytics and machine learning as a consultant. When he’s not building on AWS, you can find him fumbling around with wood projects.
Source: Read MoreÂ
