Frank Fu's Blog

OpenAI’ s API has attracted many developers with its powerful language processing capabilities, yet its pricing structure often causes confusion. Although the official documentation states that pricing is “per token”, the concept of tokens is not intuitive for most developers. A common question arises: “How much does using the OpenAI Realtime API cost per minute?”

To shed light on this, I conducted in-depth tests on two mainstream models using OpenAI Playground. With real-world data, I uncover the actual costs and provide practical optimization strategies.

1. Why Do Developers Care More About “Per Minute Call Costs”?

1.1 The Intrinsic Complexity of Token-Based Billing

OpenAI’s pricing formula seems very straightforward, it goes as input tokens + output tokens × unit price, but in real-world development, when it comes to pricing, developers often struggle with the following issues:

• Difficult cost estimation: How many tokens are needed for a 1,000-word text? How much is consumed in a 5-minute voice conversation?
• Opaque model differences: What is the token price difference between Realtime-gpt-4o-mini-preview-2024-12-17 and Realtime-gpt-4o-preview-2024-12-17?
• Hidden cost traps: How do system prompts and other fixed costs affect long-term expenses?

1.2 Real-World Business Scenarios

In applications like real-time voice customer service and intelligent assistants, per-minute costs are the primary concern:

• Budgeting: If the per-minute cost is $0.50, a 10-minute call costs only $5; if it rises to $2, the cost jumps to $20.
• Performance trade-offs: Higher-performance models offer a smoother experience but may increase costs by up to 8x (as demonstrated later).
• Scalability challenges: For 1,000 calls per day, cost differences can reach $1,500 per day, directly affecting business viability.

2. Token Mechanism: From Technical Definition to Cost Mapping

2.1 What Is a Token?

Technical definition: A token is the smallest unit of text processing, with OpenAI using the Byte Pair Encoding (BPE) algorithm for tokenization.

Example: “Hello, world!” → [“Hello”, “,”, “world”, “!”] (4 tokens)

2.2 Cost Calculation Formula

Single-call cost = (Input tokens × Input price) + (Output tokens × Output price)

For Realtime-gpt-4o-preview-2024-12-17, the pricing is:

Input price: $0.00004 per token ($40 per million tokens)

Output price: $0.00008 per token ($80 per million tokens)

Example: Processing 1,000 input tokens + 500 output tokens, the cost would be: (1000×0.00004) + (500×0.00008) = $0.08

2.3 How to Calculate Token Count?

OpenAI provides the tiktoken library to help developers precisely calculate token counts. Example:

import tiktoken

# Initialize encoder
encoder = tiktoken.get_encoding("cl100k_base")

# Calculate token count
text = "Hello, World！"
tokens = encoder.encode(text)
print(len(tokens))  # Output: 4

3. Real-World Testing: Per-Minute Costs Across Models & Configurations

3.1 Testing Methodology

To better understand OpenAI Realtime API costs, we conducted real-world tests on the OpenAI Playground. The test involved two OpenAI models:

• Realtime-gpt-4o-mini-preview-2024-12-17
• Realtime-gpt-4o-preview-2024-12-17

Each model was tested under two configurations:

• Without system prompts
• With a system prompt containing about 1,000 words (e.g., menus, conversation flows)

We extracted actual cost and token consumption data from OpenAI Playground’s Uasge, obtaining four test records.

3.2 Cost Analysis of the Economical Model: Realtime-gpt-4o-mini-preview-2024-12-17

Key insights:

• Without system prompts: Cost remains at $0.16 per minute.
• With system prompts: Fixed token cost (about 1,300 tokens) drives total cost up to $0.33 per minute.

3.3 Cost Analysis of the High-Performance Model: Realtime-gpt-4o-preview-2024-12-17

Key insights:

• Without system prompts: Cost remains at $0.18 per minute.
• With system prompts: Fixed token cost (about 1,300 tokens) drives total cost up to $1.63 per minute.

3.4 Core Findings

• Fixed costs accumulate: System prompts consume input token quotas and are billed every call.
• Model Differences: The Mini model offers better cost-effectiveness; even with about 1,300 token prompts, it costs just $0.33 per minute. The Preview model is more powerful but has a higher token price and consumes more tokens when processing long texts, requiring careful prompt usage.

4. Cost Optimization Strategies: From Code to Architecture

4.1 Three Principles for Streamlining Prompts

1）Remove redundancy:

# Before optimization: excessive description
prompt = "Welcome! We are a restaurant chain established in 2010, specializing in burgers, pizzas, and salads..."

# After optimization: concise core message
prompt = "Menu: Burger ($10), Pizza ($12), Salad ($8)"

Effect: Token count reduced from 200 to 40, cutting costs by 80%.

2）Use structured JSON instead of natural language

{
  "menu": [
    {"name": "Burger", "price": 10},
    {"name": "Pizza", "price": 12}
  ]
}

Effect: Token count reduced by 65%, improving parsing efficiency.

3）Dynamically load prompts: Call the prompt word module on demand

if user_intent == "menu_inquiry":
    load_prompt("menu_prompt.json")
elif user_intent == "customer_support":
    load_prompt("service_prompt.json")

4.2 Golden Rules for Model Selection

• Simple scenarios (FAQs, order tracking)

→ Use Realtime-gpt-4o-mini to keep costs under $0.33 per minute.

• Complex scenarios (medical consultations, multi-turn conversations)

→ Use Realtime-gpt-4o-preview, but implement staged prompt loading to keep costs under $1 per minute.

4.3 Cost-Control Strategies through Technical Means

• Enforce token limits:

response = openai.ChatCompletion.create(
    model="gpt-4o-mini",
    messages=messages,
    max_tokens=150  # Limit response length  
)

• Implement cost threshold switching:

if current_cost > $1.0/minute:
    switch_model("gpt-4o-mini")  # Auto-downgrade

5. Conclusion: Striking a Balance Between Performance and Cost

Based on our test data, the key factors in controlling OpenAI Realtime API costs are:

1）Optimizing and dynamically loading system prompts

2）Choosing the right model for the right scenario

3）Implementing real-time monitoring and cost threshold switching mechanisms

Recommendations for Developers:

• Use Mini for 80% of simple scenarios: Keep costs below $0.3/min.
• Use Preview for 20% of complex scenarios: Optimize architecture to prevent cost surges.

Through this testing, we can clearly see the significant impact of system prompts and model selection on overall costs. Developers should make informed choices based on business needs to achieve the best balance between performance and cost.

If you need further technical support or detailed test data, feel free to leave a comment!

Thanks for reading! Additionally, a video version of this blog is available below—stay tuned and enjoy watching!

And welcome to explore my Youtube channel https://www.youtube.com/@frankfu007 for more exciting content. If you enjoy my video, don’t forget to like and subscribe for more insights!

Interested in more application of Realtime API in practical projects? Explore my articles linked below!

• Education Nano 01 – Modular Wheel-Leg Robot for STEMJuly 10, 2025
• Audio-Visual Synchronization Algorithms in Digital Humans and the TIME_WAIT Challenge in WebSocket CommunicationJune 20, 2025
• Building a Voice-Controlled Robot Using OpenAI Realtime API: A Full-Link Implementation from RDK X5 to ES02June 18, 2025
• Desktop Balancing Bot(ES02)-Dual-Wheel Legged Robot with High-Performance AlgorithmJune 13, 2025
• Wheeled-Legged Robot ES01: Building with ESP32 & SimpleFOCJune 5, 2025