fix(cost): avoid double billing Gemini embedding modalities

[tin-berri]

TLDR

Problem this solves:

• Gemini embedding audio costs include both tokens and seconds
• Vertex unit tests fail after the pricing sync

How it solves it:

• Charge each modality once, preferring explicit token prices
• Preserve image/video tokens and align published online rates

User Flow

Before: 64 audio tokens representing two seconds cost $0.000736

1. POST http://localhost:45202/spend/calculate with the embedding usage below
2. Receive HTTP 200 with {"cost":0.000736}

After: the same usage costs $0.000416 at Google's token rate

1. POST http://localhost:45202/spend/calculate with the embedding usage below
2. Receive HTTP 200 with {"cost":0.000416}

Relevant issues

• Fix double billing when token and physical-unit prices coexist
• Preserve unit-only pricing and explicit zero token prices
• Triggered by chore(prices): sync Vertex AI prices: 14 models #40955; original failed Vertex job

Affected release

Linear ticket

Changes

• Select one token or physical-unit charge per modality
• Retain image/video token counts without inventing a text token
• Apply published online rates across six stable/preview aliases and synchronize the pricing schema
• Reuse existing embedding tests and add one seven-case pricing matrix

Pre-Submission checklist

• I have added meaningful tests
• The handful of test files covering my change pass locally: 390 passed
• My PR passes all required CI/CD checks
• My PR's scope is as isolated as possible; it only solves one specific problem
• Greptile 5/5 on c698972; Bugbot and Veria passed

Screenshots / Proof of Fix

Real local proxy and Postgres, using /spend/calculate with example usage and each commit's bundled prices

Save as audio-request.json:

{"completion_response":{"model":"vertex_ai/gemini-embedding-2","usage":{"prompt_tokens":64,"completion_tokens":0,"total_tokens":64,"prompt_tokens_details":{"text_tokens":0,"audio_tokens":64,"audio_length_seconds":2}}}}

Before (d1fc231)

1. Send the request

   curl -sS http://localhost:45202/spend/calculate \
     -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
     -H 'Content-Type: application/json' --data @audio-request.json

2. HTTP 200: {"cost":0.000736}

After (c698972)

1. Send the same request

   curl -sS http://localhost:45202/spend/calculate \
     -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
     -H 'Content-Type: application/json' --data @audio-request.json

2. HTTP 200: {"cost":0.000416}

Type

Bug Fix

Caveats

Low

• Paid Vertex inference unverified: Google credentials unavailable locally

Final Attestation

• Tests cover mixed units, zero rates, caching, and fallback pricing

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_vertex_embedding_audio_cost (c698972) with main (b67137b)}

[greptile-apps]

Greptile Summary

This PR prevents multimodal embedding inputs from being charged using both token and physical-unit rates.

• Selects one pricing basis per audio, image, and video modality, preferring explicitly configured token rates.
• Preserves Gemini embedding image and video token details without synthesizing text usage.
• Updates Gemini Embedding 2 pricing aliases and keeps the generated pricing schema synchronized.
• Adds regression coverage for mixed pricing units, zero rates, cached audio, modality usage conversion, and provider/model aliases.

Confidence Score: 5/5

The PR appears safe to merge; the current commit keeps the generated schema synchronized and the single-charge modality behavior is covered across relevant fallback and zero-rate cases.

No actionable new failures remain in the changes since the previous review, and no mounted repository-rule violations were identified.

Important Files Changed

Filename	Overview
litellm/litellm_core_utils/llm_cost_calc/utils.py	Introduces mutually exclusive token-or-physical-unit modality charging while preserving explicit zero rates and generic token fallback.
litellm/llms/vertex_ai/gemini_embeddings/batch_embed_content_transformation.py	Preserves image and video token metadata and removes the artificial text-token floor.
model_prices_and_context_window.json	Adds published modality token rates across the intended Gemini Embedding 2 aliases.
model_prices_and_context_window.schema.json	Adds the generated nonnegative-number schema property for video token pricing, consistent with the pricing source.
tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py	Adds a pricing matrix covering token preference, unit fallback, explicit zero rates, generic fallback, and cached audio.
tests/test_litellm/llms/vertex_ai/gemini_embeddings/test_batch_embed_content_transformation.py	Expands modality usage and exact-cost regressions across six Gemini and Vertex model/provider combinations.

_{Reviews (2): Last reviewed commit: "fix(cost): avoid double billing Gemini e..." | Re-trigger Greptile}

[tin-berri]

@greptileai Please review the current commit c698972, including the generated pricing schema and the single-charge modality regressions

[tin-berri]

bugbot run

Please review c698972 for pricing correctness, fallback behavior, cached audio accounting, and the generated schema update

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit c698972. Configure here.}