Spurious errors from litellm due to changed type of response.usage

[kouk]

• This is actually a bug report.
• I am not getting good LLM Results
• I have tried asking for help in the community on discord or discussions and have not received a response.
• I have tried searching the documentation and have not found an answer.

What Model are you using?

• gpt-3.5-turbo
• gpt-4-turbo
• gpt-4
• Other: any

Describe the bug
In #343 code was added (now living in the update_total_usage method of instructor.utils.core) that updates the response object with the total tokens used. However when used in conjunction with Litellm this introduces a bug (discussed in BerriAI/litellm#13672) because the type of response.usage unexpectedly changes from one subclass of OpenAIUsage to another (that lacks some methods that litellm added). There is code in litellm (the Langfuse integration) that depends on the response.usage field having this specific type.

To Reproduce

Run repro.py

$ uvx --with langfuse,instructor[litellm]  --exclude-newer 2026-01-01 python repro.py
python: 3.12.4 (main, Jul 12 2024, 09:00:16) [Clang 15.0.0 (clang-1500.3.9.4)]
instructor: 1.13.0
litellm: 1.80.11
langfuse: 3.11.2
14:16:02 - LiteLLM:WARNING: opentelemetry.py:186 - Proxy Server is not installed. Skipping OpenTelemetry initialization.
14:16:05 - LiteLLM:ERROR: _utils.py:273 - [Arize/Phoenix] Failed to set OpenInference span attributes: 'CompletionUsage' object has no attribute 'get'

Expected behavior
No error

[weiguangli-io]

I've submitted a fix in #2116.

The root cause is in update_total_usage(): line 350 replaces response.usage with a plain CompletionUsage object (total_usage), discarding the original litellm-specific Usage subclass. This breaks integrations like Langfuse that depend on litellm's custom methods.

The fix copies accumulated token totals back to the original usage object instead of replacing it, preserving the provider-specific type.

[kouk]

@OiPunk a fix was already implemented yesterday here #2114

[kouk]

@jxnl not sure how we ended up with 4 PRs that basically do the same thing. But in any case, can you take a look at this issue please?