fix: replace double-count multiplier with routing-prefix stripping

Refactor cost calculation logic to remove cross-region surcharge handling and strip Bedrock routing prefixes.
This commit is contained in:
Vitaliy 2026-06-20 10:11:41 -04:00 • committed by GitHub
parent 4a6ebaf183
commit a02a9eb531
No known key found for this signature in database
GPG key ID: B5690EEEBB952194

View file

@ -10,29 +10,24 @@ from litellm.litellm_core_utils.llm_cost_calc.utils import generic_cost_per_toke
if TYPE_CHECKING:
from litellm.types.utils import Usage
# AWS charges a ~10% surcharge for cross-region inference profiles (us./eu./ap. prefixes)
_CROSS_REGION_INFERENCE_SURCHARGE = 1.1
# Routing prefixes stripped before model-info lookup, ordered longest-first so
# that "bedrock/converse/" is matched before the shorter "bedrock/" prefix.
_BEDROCK_ROUTING_PREFIXES = ("bedrock/converse/", "bedrock/", "converse/")
def _is_cross_region_inference_model(model: str) -> bool:
"""Return True if *model* uses a Bedrock cross-region inference prefix.
def _strip_bedrock_routing_prefix(model: str) -> str:
"""Return *model* with any leading Bedrock routing prefix removed.
Cross-region inference profile IDs begin with a geographic abbreviation
followed by a dot, e.g. ``us.anthropic.claude-sonnet-4-6``. AWS bills
these at a ~10% premium over the equivalent base-model price.
litellm may pass model strings that still carry provider or converse routing
prefixes (e.g. ``bedrock/converse/us.anthropic.claude-sonnet-4-6``).
``get_model_info`` looks up the canonical key without these prefixes, so
stripping them first ensures the correct price entry — including cross-region
keys such as ``us.*`` — is found and used.
"""
from litellm.llms.bedrock.common_utils import (
get_bedrock_cross_region_inference_regions,
)
stripped = model
for prefix in ("bedrock/", "converse/"):
if stripped.startswith(prefix):
stripped = stripped[len(prefix) :]
break
potential_region = stripped.split(".", 1)[0]
return potential_region in get_bedrock_cross_region_inference_regions()
for prefix in _BEDROCK_ROUTING_PREFIXES:
if model.startswith(prefix):
return model[len(prefix):]
return model
def cost_per_token(
@ -43,18 +38,15 @@ def cost_per_token(
Follows the same logic as Anthropic's cost per token calculation.
For cross-region inference profiles (us./eu./ap. prefixes), applies the AWS
10% surcharge on top of the base-model token prices.
Cross-region inference model prices (``us.*``, ``eu.*``, ``ap.*``) are
already encoded in ``model_prices_and_context_window.json`` with the correct
surcharge applied — no additional multiplier is needed here. The prefix
stripping ensures that callers using ``bedrock/converse/us.*`` style strings
still resolve to the correct cross-region price entry.
"""
prompt_cost, completion_cost = generic_cost_per_token(
model=model,
return generic_cost_per_token(
model=_strip_bedrock_routing_prefix(model),
usage=usage,
custom_llm_provider="bedrock",
service_tier=service_tier,
)
if _is_cross_region_inference_model(model):
prompt_cost *= _CROSS_REGION_INFERENCE_SURCHARGE
completion_cost *= _CROSS_REGION_INFERENCE_SURCHARGE
return prompt_cost, completion_cost