test(main): pin what a streamed response costs, end to end

[yuneng-berri]

TLDR

Problem this solves:

• Streamed chunks reach a spend row through stream_chunk_builder
• No test in test_main.py asserted a dollar figure on that path
• Zeroing the reported usage there leaves all 86 green

How it solves it:

• Three cases that assert the number, not that it is positive
• Streamed and whole responses must bill the same usage the same

User Flow

No end-user behavior changes. A developer streaming a completion keeps seeing the
same spend on https://litellm-domain/ui/?page=logs, and a change that would zero
it now fails before it ships.

Relevant issues

Linear ticket

Pre-Submission checklist

• I have added meaningful tests
• The handful of test files covering my change pass locally
• My PR passes all required CI/CD checks (e.g., lint, schema.d.ts sync check, etc.)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review

Screenshots / Proof of Fix

Shared setup: the chunk builder is rewritten to discard the token counts the
provider reported in its usage chunk, one field at a time, and the mapped test
file is run against it.

in litellm/litellm_core_utils/streaming_chunk_builder_utils.py
    prompt_tokens = usage_chunk.get("prompt_tokens", 0) or 0     ->  prompt_tokens = 0
    completion_tokens = usage_chunk.get("completion_tokens", 0) or 0  ->  completion_tokens = 0

uv run pytest tests/test_litellm/test_main.py -q

Before (ff02d5c)

1. uv run pytest tests/test_litellm/test_main.py -q

86 passed, 11 warnings

2. Both rewrites, against the whole file:

SURVIVED  reported prompt tokens dropped
SURVIVED  reported completion tokens dropped

kill rate: 0/2

3. So a build that bills every streamed request at $0 ships green

After (34930e8)

1. uv run pytest tests/test_litellm/test_main.py -q

90 passed, 11 warnings in 1.13s

2. The same two rewrites:

KILLED    reported prompt tokens dropped
KILLED    reported completion tokens dropped

kill rate: 2/2

Type

✅ Test

Caveats (if any)

• The expected figure is recomputed from model_cost, not hard-coded
• A non-zero guard runs first, so all-zero prices cannot satisfy it
• The no-usage case asserts the relationship, so a tokenizer bump is fine
• Text streams only; tool calls and reasoning tokens stay unpinned

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

[greptile-apps]

Greptile Summary

This test-only PR adds end-to-end regression coverage for streamed completion cost accounting.

• Verifies provider-reported prompt and completion token usage survives stream reconstruction.
• Confirms streamed and non-streamed responses with identical usage receive the same cost.
• Covers tokenizer-based billing when a stream contains no provider-reported usage.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
tests/test_litellm/test_main.py	Adds local streamed-response cost tests without changing production behavior; no eligible follow-up finding remains.

_{Reviews (2): Last reviewed commit: "test(main): pin what a streamed response..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[yuneng-berri]

@greptileai