Usage tracking with AG2

cost calculation

1. Use AG2’s OpenAIWrapper for cost estimation

The OpenAIWrapper from autogen tracks token counts and costs of your API calls. Use the create() method to initiate requests and print_usage_summary() to retrieve a detailed usage report, including total cost and token usage for both cached and actual requests.

• mode=["actual", "total"] (default): print usage summary for non-caching completions and all completions (including cache).
• mode='actual': only print non-cached usage.
• mode='total': only print all usage (including cache).

Reset your session’s usage data with clear_usage_summary() when needed.

2. Track cost and token count for agents

We also support cost estimation for agents. Use Agent.print_usage_summary() to print the cost summary for the agent. You can retrieve usage summary in a dict using Agent.get_actual_usage() and Agent.get_total_usage(). Note that Agent.reset() will also reset the usage summary.

To gather usage data for a list of agents, we provide an utility function autogen.gather_usage_summary(agents) where you pass in a list of agents and gather the usage summary.

3. Custom token price for up-to-date cost estimation

AG2 tries to keep the token prices up-to-date. However, you can pass in a price field in config_list if the token price is not listed or up-to-date. Please creating an issue or pull request to help us keep the token prices up-to-date!

Note: in json files, the price should be a list of two floats.

Example Usage:

{
    "model": "gpt-3.5-turbo-xxxx",
    "api_key": "YOUR_API_KEY",
    "price": [0.0005, 0.0015]
}

Caution when using Azure OpenAI!

If you are using azure OpenAI, the model returned from completion doesn’t have the version information. The returned model is either ‘gpt-35-turbo’ or ‘gpt-4’. From there, we are calculating the cost based on gpt-3.5-turbo-0125: (0.0005, 0.0015) per 1k prompt and completion tokens and gpt-4-0613: (0.03, 0.06). This means the cost can be wrong if you are using a different version from azure OpenAI.

This will be improved in the future. However, the token count summary is accurate. You can use the token count to calculate the cost yourself.

Requirements

AG2 requires Python>=3.9:

pip install "ag2"

Set your API Endpoint

The config_list_from_json function loads a list of configurations from an environment variable or a json file.

import autogen
from autogen import AssistantAgent, OpenAIWrapper, UserProxyAgent, gather_usage_summary

config_list = autogen.config_list_from_json(
    "OAI_CONFIG_LIST",
    filter_dict={
        "tags": ["gpt-4o", "gpt-4o-mini"],  # comment out to get all
    },
)

It first looks for environment variable “OAI_CONFIG_LIST” which needs to be a valid json string. If that variable is not found, it then looks for a json file named “OAI_CONFIG_LIST”. It filters the configs by tags (you can filter by other keys as well).

The config list looks like the following:

config_list = [
    {
        "model": "gpt-4o",
        "api_key": "<your OpenAI API key>",
        "tags": ["gpt-4o"],
    },  # OpenAI API endpoint for gpt-4o
    {
        "model": "gpt-4o-mini",
        "base_url": "<your Azure OpenAI API base>", 
        "api_type": "azure", 
        "api_version": "2024-07-18",
        "api_key": "<your Azure OpenAI API key>",
        "tags": ["gpt-4o-mini", "20240718"],
    }
]

You can set the value of config_list in any way you prefer. Please refer to this User Guide for full code examples of the different methods.

OpenAIWrapper with cost estimation

client = OpenAIWrapper(config_list=config_list)
messages = [
    {"role": "user", "content": "Can you give me 3 useful tips on learning Python? Keep it simple and short."},
]
response = client.create(messages=messages, cache_seed=None)
print(response.cost)

OpenAIWrapper with custom token price

# Adding price to the config_list
for i in range(len(config_list)):
    config_list[i]["price"] = [
        1,
        1,
    ]  # Note: This price is just for demonstration purposes. Please replace it with the actual price of the model.

client = OpenAIWrapper(config_list=config_list)
messages = [
    {"role": "user", "content": "Can you give me 3 useful tips on learning Python? Keep it simple and short."},
]
response = client.create(messages=messages, cache_seed=None)
print("Price:", response.cost)

Usage Summary for OpenAIWrapper

When creating a instance of OpenAIWrapper, cost of all completions from the same instance is recorded. You can call print_usage_summary() to checkout your usage summary. To clear up, use clear_usage_summary().

client = OpenAIWrapper(config_list=config_list)
messages = [
    {"role": "user", "content": "Can you give me 3 useful tips on learning Python? Keep it simple and short."},
]
client.print_usage_summary()  # print usage summary

# The first creation
# By default, cache_seed is set to 41 and enabled. If you don't want to use cache, set cache_seed to None.
response = client.create(messages=messages, cache_seed=41)
client.print_usage_summary()  # default to ["actual", "total"]
client.print_usage_summary(mode="actual")  # print actual usage summary
client.print_usage_summary(mode="total")  # print total usage summary

# take out cost
print(client.actual_usage_summary)
print(client.total_usage_summary)

# Since cache is enabled, the same completion will be returned from cache, which will not incur any actual cost.
# So actual cost doesn't change but total cost doubles.
response = client.create(messages=messages, cache_seed=41)
client.print_usage_summary()

# clear usage summary
client.clear_usage_summary()
client.print_usage_summary()

# all completions are returned from cache, so no actual cost incurred.
response = client.create(messages=messages, cache_seed=41)
client.print_usage_summary()

Usage Summary for Agents

• Agent.print_usage_summary() will print the cost summary for the agent.
• Agent.get_actual_usage() and Agent.get_total_usage() will return the usage summary in a dict. When an agent doesn’t use LLM, they will return None.
• Agent.reset() will reset the usage summary.
• autogen.gather_usage_summary will gather the usage summary for a list of agents.

assistant = AssistantAgent(
    "assistant",
    system_message="You are a helpful assistant.",
    llm_config={
        "timeout": 600,
        "cache_seed": None,
        "config_list": config_list,
    },
)

ai_user_proxy = UserProxyAgent(
    name="ai_user",
    human_input_mode="NEVER",
    max_consecutive_auto_reply=1,
    code_execution_config=False,
    llm_config={
        "config_list": config_list,
    },
    # In the system message the "user" always refers to the other agent.
    system_message="You ask a user for help. You check the answer from the user and provide feedback.",
)
assistant.reset()

math_problem = "$x^3=125$. What is x?"
ai_user_proxy.initiate_chat(
    assistant,
    message=math_problem,
)

ai_user_proxy.print_usage_summary()
print()
assistant.print_usage_summary()

user_proxy = UserProxyAgent(
    name="user",
    human_input_mode="NEVER",
    max_consecutive_auto_reply=2,
    code_execution_config=False,
    default_auto_reply="That's all. Thank you.",
)
user_proxy.print_usage_summary()

print("Actual usage summary for assistant (excluding completion from cache):", assistant.get_actual_usage())
print("Total usage summary for assistant (including completion from cache):", assistant.get_total_usage())

print("Actual usage summary for ai_user_proxy:", ai_user_proxy.get_actual_usage())
print("Total usage summary for ai_user_proxy:", ai_user_proxy.get_total_usage())

print("Actual usage summary for user_proxy:", user_proxy.get_actual_usage())
print("Total usage summary for user_proxy:", user_proxy.get_total_usage())

usage_summary = gather_usage_summary([assistant, ai_user_proxy, user_proxy])
usage_summary["usage_including_cached_inference"]

Edit this page