feat(arize): emit prompt-caching tokens as OpenInference span attributes

[restato]

Relevant issues

Restores cached_tokens emission that PR #24112 added but PR #26506 inadvertently removed during a refactor for raw OpenAI Pydantic usage support.

Adds cache_creation_tokens emission, which has never been wired up despite the OpenInference LLM_TOKEN_COUNT_PROMPT_DETAILS_CACHE_WRITE constant being defined.

Pre-Submission checklist

• I have added testing in the tests/test_litellm/ directory — 4 new tests in tests/test_litellm/integrations/arize/test_arize_utils.py
• My PR passes all unit tests for the touched module (tests/test_litellm/integrations/arize/test_arize_utils.py → 8/8 passing)
• My PR's scope is as isolated as possible — single change to _set_usage_outputs to emit two additional OpenInference attributes
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Type

🐛 Bug Fix

Changes

What

_set_usage_outputs now emits prompt-cache breakdown as OpenInference span attributes:

LiteLLM normalized field	OpenInference attribute
usage.prompt_tokens_details.cached_tokens	llm.token_count.prompt_details.cache_read
usage.prompt_tokens_details.cache_creation_tokens	llm.token_count.prompt_details.cache_write

Why

LiteLLM normalizes provider-specific cache fields onto prompt_tokens_details in Usage.__init__:

• Anthropic cache_read_input_tokens → prompt_tokens_details.cached_tokens
• DeepSeek prompt_cache_hit_tokens → prompt_tokens_details.cached_tokens
• OpenAI native prompt_tokens_details.cached_tokens → already there
• Anthropic cache_creation_input_tokens → prompt_tokens_details.cache_creation_tokens

Reading from prompt_tokens_details keeps the OTEL emitter provider-agnostic.

Without these attributes, observability backends (Langfuse, Arize Phoenix) cannot display the cache breakdown nor apply correct cost calculation — Anthropic prices cache reads at 0.1× and cache writes at 1.25× standard input.

History of regression

1. PR fix(otel): use completion_tokens_details for Chat Completions API rea… #24112 — added cached_tokens emission ✅
2. PR fix(arize): _set_usage_outputs handles raw OpenAI Pydantic CompletionUsage #26506 — refactored _set_usage_outputs to handle raw OpenAI CompletionUsage Pydantic objects, but the new function body dropped the cached_tokens block. Cache reads stopped reaching observability backends. ❌
3. This PR — restores cached_tokens and adds cache_creation_tokens. ✅

Tests

tests/test_litellm/integrations/arize/test_arize_utils.py:

• test_arize_set_attributes_anthropic_cache_tokens — both cache_read and cache_write set
• test_arize_set_attributes_openai_cached_tokens — only cache_read set (no cache_write concept on OpenAI)
• test_arize_set_attributes_deepseek_cache_hit_tokens — prompt_cache_hit_tokens normalized → cache_read set
• test_arize_set_attributes_no_cache_tokens_omits_attributes — neither attribute set when usage lacks cache fields

Lint

• black --check passes on changed files
• ruff check passes on _utils.py; pre-existing T201 print warning in TestArizeLogger (line 469) is unrelated to this PR
  ~

[greptile-apps]

Greptile Summary

This PR restores the cached_tokens OpenInference span attribute that was accidentally dropped in PR #26506 and adds the previously-unwired cache_creation_tokens attribute, both emitted from _set_usage_outputs in the Arize integration. The fix reads from the provider-agnostic prompt_tokens_details field that LiteLLM already normalizes across Anthropic, DeepSeek, and OpenAI, and is covered by four new mock-only tests.

Confidence Score: 4/5

Safe to merge; the change is a small, isolated additive fix with good test coverage and no breaking changes.

Only a single P2 style finding (truthy guard vs. is not None for zero token counts). No logic bugs, no security concerns, no backwards-incompatible changes. Score capped at 4 due to the P2 finding per scoring guidelines.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/integrations/arize/_utils.py	Adds prompt-cache token emission (cache_read / cache_write) to _set_usage_outputs; dual dict/object access pattern is correct; only nit is truthy-check silencing explicit zeros.
tests/test_litellm/integrations/arize/test_arize_utils.py	Adds 4 well-scoped mock-only tests covering Anthropic, OpenAI, DeepSeek, and no-cache paths; no real network calls; assertions are precise and correct.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[_set_usage_outputs] --> B{usage present?}
    B -- No --> Z[return]
    B -- Yes --> C[emit total / completion / prompt tokens]
    C --> D[emit reasoning_tokens if present]
    D --> E[get prompt_tokens_details]
    E --> F{dict-like? hasattr get}
    F -- Yes --> G[dict .get cached_tokens / cache_creation_tokens]
    F -- No --> H[getattr cached_tokens / cache_creation_tokens]
    G --> I{cached_tokens truthy?}
    H --> I
    I -- Yes --> J[set LLM_TOKEN_COUNT_PROMPT_DETAILS_CACHE_READ]
    I -- No --> K{cache_creation_tokens truthy?}
    J --> K
    K -- Yes --> L[set LLM_TOKEN_COUNT_PROMPT_DETAILS_CACHE_WRITE]
    K -- No --> M[done]
    L --> M

Loading

_{Reviews (1): Last reviewed commit: "feat(arize): emit prompt-caching tokens ..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[restato]

@ryan-crabbe-berri Please review.

[oss-pr-review-agent-shin]

🤖 litellm-agent: This PR was marked BLOCKED 7 days ago with no subsequent activity. Closing automatically.