I can't get the token limit when using Gork with LiteLLM, while other models like GPT or Gemini work normally.

[thaisonhp]

Question

I can't get the token limit when using Gork with LiteLLM, while other models like GPT or Gemini work normally.
I had read and save the event from respone but i did't not see any info about usage of gork

[ihower]

@thaisonhp I’m not able to reproduce this issue on my side. I can see usage correctly for both Gemini and Grok when using LiteLLM.

My script:

from agents import Agent, Runner
from agents.extensions.models.litellm_model import LitellmModel

print("================================================ Gemini")

model1 = LitellmModel(model="gemini/gemini-2.5-pro")
agent1 = Agent(
    name="Assistant",
    instructions="You are a helpful assistant",
    model=model1
)
result = Runner.run_sync(agent1, "just say hello")
print(result.final_output)
print(result.context_wrapper.usage)


print("================================================ Grok")

model2 = LitellmModel(model="xai/grok-4-1-fast-reasoning")
agent2 = Agent(
    name="Assistant",
    instructions="You are a helpful assistant",
    model=model2
)

result = Runner.run_sync(agent2, "just say hello")

print(result.final_output)
print(result.context_wrapper.usage)

Output:

================================================ Gemini
Hello
Usage(requests=1, input_tokens=9, input_tokens_details=InputTokensDetails(cached_tokens=0), output_tokens=710, output_tokens_details=OutputTokensDetails(reasoning_tokens=709), total_tokens=719, request_usage_entries=[RequestUsage(input_tokens=9, output_tokens=710, total_tokens=719, input_tokens_details=InputTokensDetails(cached_tokens=0), output_tokens_details=OutputTokensDetails(reasoning_tokens=709))])
================================================ Grok
Hello!
Usage(requests=1, input_tokens=164, input_tokens_details=InputTokensDetails(cached_tokens=151), output_tokens=2, output_tokens_details=OutputTokensDetails(reasoning_tokens=72), total_tokens=238, request_usage_entries=[RequestUsage(input_tokens=164, output_tokens=2, total_tokens=238, input_tokens_details=InputTokensDetails(cached_tokens=151), output_tokens_details=OutputTokensDetails(reasoning_tokens=72))])

Also, all of these usage results are coming directly from LiteLLM. The OpenAI Agents SDK does not apply any special handling for usage parsing. So if the usage is missing, it is very likely an issue on the LiteLLM side.

[ihower]

@thaisonhp I can now reproduce your issue, and it only happens in streaming mode.

I’ve submitted a PR #2126 to add a defensive workaround on the SDK side.

I also believe this is primarily a LiteLLM issue, and I’ve reported it here: BerriAI/litellm#17136