Large Language Models
Large language models (LLM) are AI systems capable of understanding and generating human language by processing vast amounts of text data.
They are deep-learning models pre-trained on enormous amounts of data, allowing them to learn complex patterns, perform analyses, and generate human-like responses.
Moreover, their ability to analyze and produce complex natural language makes them invaluable for businesses and developers across multiple industries.
In recent years, Large Language Model has taken the limelight of various tech giants as well as innovative startups. All striving to deliver the best possible results through their respective LLMs.
As these models become more sophisticated, the competition has intensified, with each player seeking to outdo the others by improving accuracy, scalability, and accessibility.
With so many options available in the market with different abilities and prices With so many options available in the market with different abilities and prices.
It has been an overhead for consumers to calculate the overall cost of utilizing the LLM model in their project.
As LLMs become more prevalent across various industries, understanding the cost structure behind them is crucial for businesses looking to integrate these powerful AI tools into their operations.
Factors Affecting LLM Price Estimation
LLM price calculation is based on different providers, model selection, input tokens, and generated output tokens.
Tokens
A token is a unit of text that is used to represent a word, phrase, or other piece of text. For example, “What is Bagisto?” in this line there are 4 tokens “What,” “is”, “Bagisto” and “?”.
Providers set the prices on input tokens and output tokens. Moreover, Every LLM have a different tokenization algorithms, thus token count for the same input may vary for different LLMs.
Models and pricing
1 Hosted LLM models:-
There are many LLM providers like OpenAI, Gemini, Groq, Openrouter, Cerebras, Anthropic, Qwen Llama, etc., who provide APIs for their cloud hosted models at different costs.
These models differ in Context length, parameters and their ability to understand and generate responses and thus price according to their ability.
Here is the list of some hosted LLM models and their price per million token :
| MODELS | API Provider | INPUT PRICE PER MILLION TOKEN | OUTPUT PRICE PER MILLION TOKEN |
| Gemini-2.5-flash | Google | $0.30 | $2.50 |
| GPT OSS 120B | Cerebras | $0.35 | $0.75 |
| GPT OSS 120B | Groq | $0.15 | $0.75 |
| GPT OSS 120B | Openrouter | $0.05 | $0.25 |
| GPT-4.1 mini | Openai | $0.40 | $1.60 |
| Qwen 3 235B | Cerebras | $0.60 | $1.20 |
| GPT-5 mini | Openai | $0.25 | $2.00 |
| GPT-5 | Openai | $1.25 | $10.00 |
| Gemini-2.5-pro | Google | $2.50 | $15.00 |
| GPT OSS 20B | Openrouter | $0.04 | $0.15 |
| Claude Sonnet 3.7 | Anthropic | $3.00 | $15.00 |
| GPT OSS 20B | Groq | $0.10 | $0.50 |
How to Estimate Price for a query in RAG Chatbot
1) Price Calculation for Input
Step-1 : Calculate the number of Token in a query, For example, for a simple query like “show me some t-shirts”. We have utilized 5 tokens i.e – “show”, “me”, “some”, “t”, “-shirts”.
Step-2 : Moreover, calculate the number of Token utilized in prompt. In our product chatbot, the prompt uses approximately 200 tokens.
Step 3 : Calculate the token consumed by the retrieved document/data from vector DB for RAG. Roughly, around 2000 tokens were consumed by data of 12 products retrieved from vector database.
Step 4 : Calculate the cost of input for the above query by using formula below:-
Input Cost = ((query tokens + prompt tokens + document/data tokens) * Input price of model)/1,000,000
For example: If Gemini-2.5-flash model is used, then the cost for input for above query to the RAG chatbot will be around $0.00066
2) Price Calculation for Output
After LLM generates the response, you can calculate the output tokens. Our chatbot consumed around 2000 tokens for the above query. Now you can calculate the output using formula:-
Output Cost = (output tokens * output price of model)/1,000,000
For example: If Gemini-2.5-flash model is used, then the cost of output for the response generated by chatbot will be around $0.00450
3) Total Price Calculation
Total Cost = Input Cost + Output Cost
Total cost for the query can be calculated by adding input cost and output cost. Thus, for the above query, “show me some t-shirts” the price for Gemini-2.5-flash model will be $0.00516
NOTE: This is the estimated cost for the above query; however, it can vary depending on the query itself, the data retrieved, and the response generated by the LLM.
Here is the list of prices for different LLMs based on the above query example
| MODELS | API Provider | ESTIMATED INPUT PRICE PER QUERY | ESTIMATED OUTPUT PRICE PER QUERY | ESTIMATED TOTAL PRICE PER QUERY |
| Gemini-2.5-flash | Google | $0.00066 | $0.00450 | $0.00516 |
| GPT OSS 120B | Cerebras | $0.00077 | $0.00135 | $0.00212 |
| GPT OSS 120B | Groq | $0.00033 | $0.00135 | $0.00168 |
| GPT OSS 120B | Openrouter | $0.00011 | $0.00045 | $0.00056 |
| GPT-4.1 mini | Openai | $0.00088 | $0.00288 | $0.00376 |
| Qwen 3 235B | Cerebras | $0.00132 | $0.00216 | $0.00348 |
| GPT-5 mini | Openai | $0.00055 | $0.00360 | $0.00415 |
| GPT-5 | Openai | $0.00275 | $0.01800 | $0.02075 |
| Gemini-2.5-pro | Google | $0.00550 | $0.02700 | $0.03250 |
| GPT OSS 20B | Openrouter | $0.00009 | $0.00027 | $0.00036 |
| Claude Sonnet 3.7 | Anthropic | $0.06600 | $0.02700 | $0.09300 |
| GPT OSS 20B | Groq | $0.00022 | $0.00090 | $0.00112 |
Check our RAG based chatbot for Magento e-commerce platform.
2 Locally hosted LLM models:-
There are various open source Large Language Models that can be deployed locally on a GPU instance that is wholly available for you only.
Moreover, these Models have more privacy than hosted models as your data is not sent to a third party models, and it is not subject to their terms of service.
The cost of these models do not depend on the number of hits and tokens used, but on the size of the GPU instance required to host the LLM model. The exact requirements depend on the model size.
a) Small to Medium LLMs (less than 8B parameters)
There are various small open source models like Owen 1.7 B/4B / 8B models, Llama 3.1 8B etc. These models can be hosted with instance having configuration given below :
- CPU: 8 cores (x86_64 or ARM)
- RAM: 16 GB minimum (32 GB recommended)
- GPU: NVIDIA GPU with 8–12 GB VRAM for acceleration
- Storage: 30–50 GB SSD free space
- OS: Linux
Some recommended hosting below :
| INSTANCE NAME | PLATFORM | STARTING PRICE PER HOUR |
| NVIDIA T4 (16GB) PCIe | RunPod | $1.50 |
| g4ad.xlarge | AWS | $0.38 |
| On-demand 1x NVIDIA Quadro RTX 6000 | lambda.ai | $0.50 |
b) Large LLMs (13B–30B parameters)
Models like Qwen3 14B / 30B, gpt-oss 20B are large open source models that can perform complex task with accuracy.
Additionally, these models can be hosted with instance having configuration given below :
- CPU: 16 + cores (x86_64 or ARM)
- RAM: 64 GB minimum (128 GB recommended)
- GPU: NVIDIA RTX 3090/4090 or A100 with 24–40 GB VRAM
- Storage: 100–200 GB SSD free space
- OS: Linux
Note: Multi-GPU support required for 30B+ models
Some recommended hosting below :
| INSTANCE NAME | PLATFORM | STARTING PRICE PER HOUR |
| g4ad.16xlarge | AWS | $3.47 |
| NVIDIA T4 (16GB) PCIe *4 | RunPod | $5.09 |
| On-demand 1x NVIDIA H100 PCIe | Vast.ai | $2.49 |
Conclusion
Incorporating Large Language Models (LLMs) into your projects requires a thorough understanding of the cost structure.
Additionally, various factors, such as the number of tokens, the selected model, and its context length, play a significant role in determining the price.
By carefully calculating token usage and selecting the appropriate model.
Moreover, businesses can optimize their LLM costs, making it easier to integrate these powerful tools into their operations without incurring excessive expenses.
Understanding these factors will enable you to make informed decisions about which model to choose based on your specific needs and budget constraints.
Be the first to comment.