[Bug]: OpenAI Responses response service_tier=default is billed as priority when requested priority

[silencedoctor]

What happened?

OpenAI Responses cost tracking can bill a request at priority rates even when the provider response says the request was served as default.

The problematic case is:

• request asks for service_tier="priority"
• OpenAI Responses response contains service_tier="default"
• LiteLLM cost calculation currently prefers the request-level tier before the response tier

That means cost tracking can select input_cost_per_token_priority / output_cost_per_token_priority even though the served tier reported by the provider was default.

Expected behavior: when a provider reports the served tier on the response (or usage), cost tracking should prefer that concrete served tier over the request preference. A response-level default should map to standard pricing.

Steps to reproduce

import litellm
from litellm import completion_cost
from litellm.types.llms.openai import ResponsesAPIResponse

model = "test-openai-responses-default-tier-cost-model"
litellm.register_model(
    model_cost={
        model: {
            "input_cost_per_token": 0.001,
            "output_cost_per_token": 0.002,
            "input_cost_per_token_priority": 0.002,
            "output_cost_per_token_priority": 0.004,
            "litellm_provider": "openai",
            "max_tokens": 8192,
            "mode": "responses",
        }
    }
)

usage = {"input_tokens": 1000, "output_tokens": 500, "total_tokens": 1500}
response = ResponsesAPIResponse(
    id="resp_default_tier",
    created_at=0,
    model=model,
    object="response",
    output=[],
    usage=usage,
    service_tier="default",
)

cost = completion_cost(
    completion_response=response,
    model=model,
    custom_llm_provider="openai",
    optional_params={"service_tier": "priority"},
    service_tier="priority",
)

Observed on current main: cost uses priority pricing.

Expected: cost uses standard/default pricing because the response says the served tier was default.

Relevant log output

In the reproduction above:

standard cost: 2.0
priority cost: 4.0
current cost with request=priority and response=default: 4.0
expected cost: 2.0

What part of LiteLLM is this about?

SDK (litellm Python package)

What LiteLLM version are you on?

Reproduced on current main (88e03e548716a45284597edf2b7f47a7e6a66d5f).