"""COMPASS cost computation utilities"""
LLM_COST_REGISTRY = {
"o1": {"prompt": 15, "response": 60},
"o3-mini": {"prompt": 1.1, "response": 4.4},
"gpt-4.5": {"prompt": 75, "response": 150},
"gpt-4o": {"prompt": 2.5, "response": 10},
"gpt-4o-mini": {"prompt": 0.15, "response": 0.6},
"gpt-4.1": {"prompt": 2, "response": 8},
"gpt-4.1-mini": {"prompt": 0.4, "response": 1.6},
"gpt-4.1-nano": {"prompt": 0.1, "response": 0.4},
"gpt-5": {"prompt": 1.25, "response": 10},
"gpt-5-mini": {"prompt": 0.25, "response": 2},
"gpt-5-nano": {"prompt": 0.05, "response": 0.4},
"gpt-5-chat-latest": {"prompt": 1.25, "response": 10},
"compassop-gpt-4o": {"prompt": 2.5, "response": 10},
"compassop-gpt-4o-mini": {"prompt": 0.15, "response": 0.6},
"compassop-gpt-4.1": {"prompt": 2, "response": 8},
"compassop-gpt-4.1-mini": {"prompt": 0.4, "response": 1.6},
"compassop-gpt-4.1-nano": {"prompt": 0.1, "response": 0.4},
"compassop-gpt-5": {"prompt": 1.25, "response": 10},
"compassop-gpt-5-mini": {"prompt": 0.25, "response": 2},
"compassop-gpt-5-nano": {"prompt": 0.05, "response": 0.4},
"compassop-gpt-5-chat-latest": {"prompt": 1.25, "response": 10},
"egswaterord-gpt4.1-mini": {"prompt": 0.4, "response": 1.6},
"wetosa-gpt-4o": {"prompt": 2.5, "response": 10},
"wetosa-gpt-4o-mini": {"prompt": 0.15, "response": 0.6},
"wetosa-gpt-4.1": {"prompt": 2, "response": 8},
"wetosa-gpt-4.1-mini": {"prompt": 0.4, "response": 1.6},
"wetosa-gpt-4.1-nano": {"prompt": 0.1, "response": 0.4},
"wetosa-gpt-5": {"prompt": 1.25, "response": 10},
"wetosa-gpt-5-mini": {"prompt": 0.25, "response": 2},
"wetosa-gpt-5-nano": {"prompt": 0.05, "response": 0.4},
"wetosa-gpt-5-chat-latest": {"prompt": 1.25, "response": 10},
"text-embedding-ada-002": {"prompt": 0.10},
}
"""LLM Costs registry
The registry maps model names to a dictionary that contains the cost
(in $/million tokens) for both prompt and response tokens.
"""
[docs]
def cost_for_model(model_name, prompt_tokens, completion_tokens):
"""Compute the API costs for a model given the token usage
Parameters
----------
model_name : str
Name of the model. Needs to be registered as a key in
:obj:`LLM_COST_REGISTRY` for this method to return a non-zero
value.
prompt_tokens, completion_tokens : int
Number of prompt and completion tokens used, respectively.
Returns
-------
float
Total cost based on the token usage.
"""
model_costs = LLM_COST_REGISTRY.get(model_name, {})
prompt_cost = prompt_tokens / 1e6 * model_costs.get("prompt", 0)
response_cost = completion_tokens / 1e6 * model_costs.get("response", 0)
return prompt_cost + response_cost
[docs]
def compute_cost_from_totals(totals):
"""Compute total cost from total tracked usage
Parameters
----------
totals : dict
Dictionary where keys are model names and their corresponding
usage statistics are values. Each usage statistics dictionary
should contain "prompt_tokens" and "response_tokens" keys
indicating the number of tokens used for prompts and responses,
respectively. This dictionary is typically obtained from the
`tracker_totals` property of a
:class:`compass.services.usage.UsageTracker` instance.
Returns
-------
float
Total cost based on the tracked usage.
"""
return sum(
cost_for_model(
model,
prompt_tokens=usage.get("prompt_tokens", 0),
completion_tokens=usage.get("response_tokens", 0),
)
for model, usage in totals.items()
)
[docs]
def compute_total_tokens_from_totals(totals):
"""Compute total prompt/response token counts from tracked usage
Parameters
----------
totals : dict
Same shape as :func:`compute_cost_from_totals`: maps model name
to a dict with ``"prompt_tokens"`` and ``"response_tokens"``.
Returns
-------
dict
``{"prompt_tokens": int, "response_tokens": int}`` summed across
all models in ``totals``.
"""
return {
"prompt_tokens": sum(
u.get("prompt_tokens", 0) for u in totals.values()
),
"response_tokens": sum(
u.get("response_tokens", 0) for u in totals.values()
),
}
[docs]
def compute_total_cost_and_token_from_totals(totals):
"""Compute total cost and token counts together from tracked usage
Returns
-------
dict
Keys: ``"cost"`` (float), ``"prompt_tokens"`` (int),
``"response_tokens"`` (int).
"""
return {
"cost": compute_cost_from_totals(totals),
**compute_total_tokens_from_totals(totals),
}
[docs]
def compute_total_cost_from_usage(tracked_usage):
"""Compute total cost from total tracked usage
Parameters
----------
tracked_usage : compass.services.usage.UsageTracker or dict
Dictionary where keys are usage categories (typically
jurisdiction names) and values are dictionaries containing usage
details. The usage details dictionaries should have a
"tracker_totals" key, which maps to another dictionary. This
innermost dictionary should have model names as keys and their
corresponding usage statistics as values. Each usage statistics
dictionary should contain "prompt_tokens" and "response_tokens"
keys indicating the number of tokens used for prompts and
responses, respectively.
Returns
-------
float
Total LLM cost based on the tracked usage.
"""
return sum(
compute_cost_from_totals(usage.get("tracker_totals", {}))
for usage in tracked_usage.values()
)
