From d793d2917fe87f2c7a45346ce5e79d29258dec06 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Wed, 28 Oct 2026 19:07:38 -0700
Subject: [PATCH] fix(proxy): read only camelCase Gemini output fields in the
 budget hold

/v1beta forwards only generationConfig.maxOutputTokens to non-Gemini models, so a snake_case cap the model never receives must not shrink the hold. Snake_case bodies now get the default output hold.
---
 litellm/proxy/spend_tracking/budget_reservation.py  | 13 ++-----------
 .../proxy/spend_tracking/test_budget_reservation.py |  2 +-
 2 files changed, 3 insertions(+), 12 deletions(-)

diff --git a/litellm/proxy/spend_tracking/budget_reservation.py b/litellm/proxy/spend_tracking/budget_reservation.py
index d20e2371f09..5cada93bd33 100644
--- a/litellm/proxy/spend_tracking/budget_reservation.py
+++ b/litellm/proxy/spend_tracking/budget_reservation.py
@@ -7,7 +7,6 @@ import time
 from collections.abc import AsyncIterator, Mapping, Sequence
 from dataclasses import dataclass
 from datetime import datetime, timedelta, timezone
-from itertools import product
 from types import MappingProxyType
 from typing import Final, NoReturn, SupportsFloat, SupportsIndex, SupportsInt, cast
 
@@ -1652,7 +1651,6 @@ def _estimate_output_tokens(
 
 
 _OUTPUT_TOKEN_FIELDS: Final = ("max_completion_tokens", "max_tokens", "max_output_tokens")
-_GEMINI_CONFIG_FIELDS: Final = ("generationConfig", "generation_config")
 
 
 def _nested_field(request_body: Mapping[str, object], config_field: str, field: str) -> object:
@@ -1660,18 +1658,11 @@ def _nested_field(request_body: Mapping[str, object], config_field: str, field:
     return config.get(field) if config is not None else None
 
 
-def _gemini_config_values(request_body: Mapping[str, object], *fields: str) -> tuple[object, ...]:
-    return tuple(
-        _nested_field(request_body, config_field, field)
-        for config_field, field in product(_GEMINI_CONFIG_FIELDS, fields)
-    )
-
-
 def _requested_output_tokens(request_body: Mapping[str, object]) -> int | None:
     candidates: Final = (
         *(request_body.get(field) for field in _OUTPUT_TOKEN_FIELDS),
         _nested_field(request_body, "inferenceConfig", "maxTokens"),
-        *_gemini_config_values(request_body, "maxOutputTokens", "max_output_tokens"),
+        _nested_field(request_body, "generationConfig", "maxOutputTokens"),
     )
     return next((tokens for tokens in map(_to_int, candidates) if tokens is not None), None)
 
@@ -1680,7 +1671,7 @@ def _get_output_multiplier(request_body: Mapping[str, object]) -> int:
     candidates: Final = (
         request_body.get("n"),
         request_body.get("best_of"),
-        *_gemini_config_values(request_body, "candidateCount", "candidate_count"),
+        _nested_field(request_body, "generationConfig", "candidateCount"),
     )
     return max((1, *(count for count in map(_to_int, candidates) if count is not None)))
 
diff --git a/tests/unit/proxy/spend_tracking/test_budget_reservation.py b/tests/unit/proxy/spend_tracking/test_budget_reservation.py
index 2b654d9085d..97f65a4402d 100644
--- a/tests/unit/proxy/spend_tracking/test_budget_reservation.py
+++ b/tests/unit/proxy/spend_tracking/test_budget_reservation.py
@@ -193,7 +193,7 @@ GEMINI_SYSTEM_PARTS: Final = ("Answer tersely", "Use British spelling")
     (
         ({"generationConfig": {"maxOutputTokens": 5}}, {"max_tokens": 5}),
         ({"generationConfig": {"maxOutputTokens": 5, "candidateCount": 3}}, {"max_tokens": 5, "n": 3}),
-        ({"generation_config": {"max_output_tokens": 5, "candidate_count": 3}}, {"max_tokens": 5, "n": 3}),
+        ({"generation_config": {"max_output_tokens": 5, "candidate_count": 3}}, {}),
         ({}, {}),
     ),
 )