fix(databricks): delegate to generic_cost_per_token for cache/audio/reasoning

[michaelxer]

TLDR

Problem this solves:

• Databricks cost calc ignored cache/audio/reasoning usage fields
• Hand-rolled math billed every prompt token at full input rate
• Same class of bug Fireworks hit before fix(fireworks_ai): bill prompt-cache hits at cache_read rate #33714

How it solves it:

• Keep Databricks model alias remapping
• Delegate token billing to generic_cost_per_token
• Aligns Databricks with DeepSeek / XAI / other OpenAI-compatible providers

Relevant issues

Fixes #35608

Linear ticket

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• 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 (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

Screenshots / Proof of Fix

Local unit verification (no mocks of the cost math itself; synthetic Usage objects exercise the real generic path):

python -m pytest tests/test_litellm/llms/databricks/test_databricks_cost_calculator.py -q -p no:inline_snapshot
....                                                                     [100%]
4 passed in 0.67s

python -m pytest tests/test_litellm/llms/databricks/test_databricks_pricing.py -q -p no:inline_snapshot
.                                                                        [100%]
1 passed in 0.52s

What the new tests pin:

1. With a cache-read rate present, cached prompt tokens bill at cache_read_input_token_cost instead of the full input rate (the old path would have charged full rate for every token)
2. Zero cached tokens still matches the full input rate
3. A warm (mostly-cached) call is cheaper than a cold call on the same prompt size
4. Legacy foundation-model aliases still remap through _resolve_databricks_base_model

I do not have Databricks credentials on this machine, so I could not hit a live Databricks endpoint for an e2e spend proof. Happy to re-run against a live proxy if a maintainer wants that.

Type

🐛 Bug Fix

Changes

litellm/llms/databricks/cost_calculator.py used to do:

prompt_cost = usage["prompt_tokens"] * model_info["input_cost_per_token"]
completion_cost = usage["completion_tokens"] * model_info["output_cost_per_token"]

That ignores prompt_tokens_details.cached_tokens, cache-creation tokens, audio tokens, and reasoning tokens. Issue #35608 called this out as the same shape as the Fireworks bug fixed in #33714.

This PR keeps the existing Databricks deployment-name remapping (dbrx / llama / mixtral / embeddings aliases), then hands billing to generic_cost_per_token(..., custom_llm_provider="databricks"), which already handles those Usage fields for DeepSeek, XAI, and other OpenAI-compatible providers.

Today no Databricks pricing-table entry publishes cache_read_input_token_cost, so there is no active mis-billing in production numbers. The risk is latent: the moment a Databricks entry gains a cache-read rate (as Fireworks did), cached tokens would silently bill at the full input rate without this change. The regression tests inject a temporary cache-read rate so that path is covered now.

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

[codecov]

Codecov Report

❌ Patch coverage is 68.42105% with 6 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
litellm/llms/databricks/cost_calculator.py	68.42%	6 Missing ⚠️

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing michaelxer:michaelxer/fix-databricks-generic-cost-20260803 (6fcf7e8) with litellm_internal_staging (e64536c)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (e2950a8) during the generation of this report, so c93a336 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[michaelxer]

Fixed the lint failure on this branch: swapped typing.Tuple for native tuple[] in cost_calculator.py so the strict UP006/UP035 budget stays clean. Focused databricks cost tests still pass locally (4/4).

[michaelxer]

Closing this draft — #37975 already landed the same Databricks cost path on staging (generic_cost_per_token plus cache rates). This branch is also conflicting and a long way behind, so there is nothing left to land here.