Completion Token Usage & Cost

By default LiteLLM returns token usage in all completion requests (See here)

However, we also expose 5 helper functions + [NEW] an API to calculate token usage across providers:

• encode: This encodes the text passed in, using the model-specific tokenizer. Jump to code

• decode: This decodes the tokens passed in, using the model-specific tokenizer. Jump to code

• token_counter: This returns the number of tokens for a given input - it uses the tokenizer based on the model, and defaults to tiktoken if no model-specific tokenizer is available. Jump to code

• cost_per_token: This returns the cost (in USD) for prompt (input) and completion (output) tokens. Uses the live list from api.litellm.ai. Jump to code

• completion_cost: This returns the overall cost (in USD) for a given LLM API Call. It combines token_counter and cost_per_token to return the cost for that query (counting both cost of input and output). Jump to code

• get_max_tokens: This returns a dictionary for a specific model, with it's max_tokens, input_cost_per_token and output_cost_per_token. Jump to code

• model_cost: This returns a dictionary for all models, with their max_tokens, input_cost_per_token and output_cost_per_token. It uses the api.litellm.ai call shown below. Jump to code

• register_model: This registers new / overrides existing models (and their pricing details) in the model cost dictionary. Jump to code

• api.litellm.ai: Live token + price count across all supported models. Jump to code

📣 This is a community maintained list. Contributions are welcome! ❤️

Example Usage

1. encode

Encoding has model-specific tokenizers for anthropic, cohere, llama2 and openai. If an unsupported model is passed in, it'll default to using tiktoken (openai's tokenizer).

from litellm import encode, decode


def test_encoding_and_decoding():
    try: 
        sample_text = "Hellö World, this is my input string!"

        # openai tokenizer 
        openai_tokens = token_counter(model="gpt-3.5-turbo", text=sample_text)

        openai_text = decode(model="gpt-3.5-turbo", tokens=openai_tokens)

        assert openai_text == sample_text
    except: 
        pass

test_encoding_and_decoding()

2. decode

Decoding is supported for anthropic, cohere, llama2 and openai.

from litellm import encode, decode


def test_encoding_and_decoding():
    try: 
        sample_text = "Hellö World, this is my input string!"

        # openai tokenizer 
        openai_tokens = token_counter(model="gpt-3.5-turbo", text=sample_text)

        openai_text = decode(model="gpt-3.5-turbo", tokens=openai_tokens)

        assert openai_text == sample_text
    except: 
        pass

test_encoding_and_decoding()

3. token_counter

from litellm import token_counter

messages = [{"user": "role", "content": "Hey, how's it going"}]
print(token_counter(model="gpt-3.5-turbo", messages=messages))

4. cost_per_token

from litellm import cost_per_token

prompt_tokens =  5
completion_tokens = 10
prompt_tokens_cost_usd_dollar, completion_tokens_cost_usd_dollar = cost_per_token(model="gpt-3.5-turbo", prompt_tokens=prompt_tokens, completion_tokens=completion_tokens))

print(prompt_tokens_cost_usd_dollar, completion_tokens_cost_usd_dollar)

5. completion_cost

• Input: Accepts a litellm.completion() response OR prompt + completion strings
• Output: Returns a float of cost for the completion call

litellm.completion()

from litellm import completion, completion_cost

response = completion(
            model="bedrock/anthropic.claude-v2",
            messages=messages,
            request_timeout=200,
        )
# pass your response from completion to completion_cost
cost = completion_cost(completion_response=response)
formatted_string = f"${float(cost):.10f}"
print(formatted_string)

prompt + completion string

from litellm import completion_cost
cost = completion_cost(model="bedrock/anthropic.claude-v2", prompt="Hey!", completion="How's it going?")
formatted_string = f"${float(cost):.10f}"
print(formatted_string)

6. get_max_tokens

• Input: Accepts a model name - e.g. gpt-3.5-turbo (to get a complete list, call litellm.model_list)
• Output: Returns a dict object containing the max_tokens, input_cost_per_token, output_cost_per_token

from litellm import get_max_tokens 

model = "gpt-3.5-turbo"

print(get_max_tokens(model)) # {'max_tokens': 4000, 'input_cost_per_token': 1.5e-06, 'output_cost_per_token': 2e-06}

7. model_cost

• Output: Returns a dict object containing the max_tokens, input_cost_per_token, output_cost_per_token for all models on community-maintained list

from litellm import model_cost 

print(model_cost) # {'gpt-3.5-turbo': {'max_tokens': 4000, 'input_cost_per_token': 1.5e-06, 'output_cost_per_token': 2e-06}, ...}

8. register_model

• Output: Returns updated model_cost dictionary + updates litellm.model_cost with model details.

from litellm import register_model

litellm.register_model({
        "gpt-4": {
        "max_tokens": 8192, 
        "input_cost_per_token": 0.00002, 
        "output_cost_per_token": 0.00006, 
        "litellm_provider": "openai", 
        "mode": "chat"
    },
})

9. api.litellm.ai

Example Curl Request

curl 'https://api.litellm.ai/get_max_tokens?model=claude-2'

{
    "input_cost_per_token": 1.102e-05,
    "max_tokens": 100000,
    "model": "claude-2",
    "output_cost_per_token": 3.268e-05
}