Explore Azure Open AI Pricing Options 

Like Azure’s services, such as Azure Cognitive Services for AI, Azure OpenAI’s cost operates on a pay-as-you-go consumption approach. It ensures you pay solely for the resources you utilize. The price per unit is determined by the model’s type and size and the quantity of tokens employed in both input and output. 

GPT-3.5 models 

The current gpt-3.5-turbo has a 4,096-token limit, while the newer gpt-3.5-turbo-16k has a 16,384-token limit. Both are priced at $0.002 per 1,000 tokens for prompts or completions. 

GPT-4 models 

GPT-4 models come in two versions: the GPT -4 model, which has an 8,192 token limit, and the GPT-4 32k, which has a 32,768 token limit.  

The GPT-4 model in prompt mode costs $0.03 for an 8K context and $0.06 for a 32K context per 1,000 tokens. In completion mode, the prices are $0.06 for an 8K context and $0.12 for a 32K context per 1,000 tokens. 

Tokens 

Tokens can be visualized as fragments of words. Before the API handles prompts, the input is segmented into tokens. These tokens may encompass trailing spaces and even parts of words. Here are some valuable approximations to understand token lengths:  

• 1 token is equal to 4 characters in English.  

• 1 token is approximately ¾ of a word.  
• 100 tokens amount to around 75 words.  

Alternatively,  

• 1–2 sentences equate to about 30 tokens.  
• 1 paragraph comprises roughly 100 tokens.  
• 1,500 words correspond to approximately 2048 tokens.  

How words are divided into tokens depends on the language. For instance, the phrase ‘Cómo estás’ (‘How are you’ in Spanish) consists of 5 tokens (covering 10 characters). The higher ratio of tokens to characters can increase the cost of using the API for languages other than English.  

To explore tokenization further, you can utilize our interactive Tokenizer tool to calculate token numbers and observe how text is segmented. If you prefer programmatically tokenizing text, consider using Tiktoken, a rapid BPE tokenizer designed for OpenAI models. Other libraries, like the transformers package for Python or the GPT-3 encoder package for node.js, are also worth exploring.  

Depending on the chosen model, requests can use up to 4,097 tokens shared between prompts and completions. If your prompt uses 4,000 tokens, your completion can have a maximum of 97 tokens. 

Fine-tuned model 

You can only fine-tune GPT-3 models (Ada, curie, DaVinci, Babbage), collectively known as “base” models.  

According to Microsoft Learn, the charges for Azure OpenAI fine-tuned models depend on three factors:  

1. Training hours  
2. Hosting hours  
3. Inference per 1,000 tokens  

Consider the hosting hours cost carefully. Once a fine-tuned model is deployed, it accumulates an hourly cost, regardless of its active usage. Monitoring the costs of fine-tuned models is essential. 

*** Currently, Azure OpenAI Service does not support fine-tuned models. 

Embedding model 

Azure OpenAI Service provides an embedding model alongside image and large language models. The standard embedding model, Ada, costs $0.0001 per 1,000 tokens. 

DALL-E 

Azure OpenAI Service, including image models, operates with pricing determined by the number of images processed, making it relevant for those exploring “Azure OpenAI Service pricing.” The standard image model, DALL-E, costs $2 for every 100 images.