From 70e3f1653f9180a613a2a58a63b9e1b102dcbf58 Mon Sep 17 00:00:00 2001
From: Anuj7411 <ojhaanuj224@gmail.com>
Date: Wed, 26 Aug 2026 19:06:44 +0530
Subject: [PATCH] fix(cost): bill non-Vertex TTS calls on raw character count,
 not whitespace-stripped

completion_cost() stripped whitespace from the billed character count for
every character-priced text-to-speech call, regardless of provider. That
convention is specific to Vertex AI's character-counting rules; OpenAI,
Azure, ElevenLabs, Groq, Minimax, and AWS Polly all bill on the raw string
length. This undercounted cost for any TTS input containing spaces, which
is effectively every real request.

Gate the whitespace-stripping to the vertex_ai provider only, matching the
same gate already used for chat/completion cost a few lines below.
---
 litellm/cost_calculator.py                  |  8 ++++-
 tests/local_testing/test_completion_cost.py | 33 +++++++++++++++++++++
 2 files changed, 40 insertions(+), 1 deletion(-)

diff --git a/litellm/cost_calculator.py b/litellm/cost_calculator.py
index 8f7cd09d364..27b5da54224 100644
--- a/litellm/cost_calculator.py
+++ b/litellm/cost_calculator.py
@@ -1411,7 +1411,13 @@ def completion_cost(
                         video_resolution=video_resolution,
                     )
                 elif call_type in _SPEECH_CALL_TYPES:
-                    prompt_characters = litellm.utils._count_characters(text=prompt)
+                    # Vertex AI bills TTS characters excluding whitespace; every other
+                    # character-priced TTS provider (OpenAI, Azure, ElevenLabs, Groq,
+                    # Minimax, AWS Polly) bills the raw character count.
+                    if custom_llm_provider == "vertex_ai":
+                        prompt_characters = litellm.utils._count_characters(text=prompt)
+                    else:
+                        prompt_characters = len(prompt)
                 elif call_type in _TRANSCRIPTION_CALL_TYPES:
                     # Check _hidden_params first (duration stored there to
                     # avoid polluting the response body), then fall back to
diff --git a/tests/local_testing/test_completion_cost.py b/tests/local_testing/test_completion_cost.py
index f47b40f2ef1..f3a9d467ae8 100644
--- a/tests/local_testing/test_completion_cost.py
+++ b/tests/local_testing/test_completion_cost.py
@@ -863,6 +863,39 @@ def test_completion_cost_tts(model):
     assert cost > 0
 
 
+@pytest.mark.parametrize("model", ["tts-1", "azure/tts-1"])
+def test_completion_cost_tts_bills_whitespace_characters(model):
+    """
+    Non-Vertex TTS providers (OpenAI, Azure, ElevenLabs, Groq, Minimax, AWS Polly)
+    bill on the raw character count, including whitespace. Only Vertex AI's
+    character-counting convention excludes whitespace. Regression test for a bug
+    where every TTS provider incorrectly had whitespace stripped before billing,
+    undercounting cost for inputs containing spaces.
+    """
+    os.environ["LITELLM_LOCAL_MODEL_COST_MAP"] = "True"
+    litellm.model_cost = litellm.get_model_cost_map(url="")
+
+    prompt_with_whitespace = "a b c d e"  # 9 chars, 4 of them whitespace
+    prompt_without_whitespace = "abcde"  # 5 chars, matches non-whitespace count above
+
+    cost_with_whitespace = completion_cost(
+        model=model,
+        prompt=prompt_with_whitespace,
+        call_type="speech",
+    )
+    cost_without_whitespace = completion_cost(
+        model=model,
+        prompt=prompt_without_whitespace,
+        call_type="speech",
+    )
+
+    input_cost_per_character = litellm.model_cost[model]["input_cost_per_character"]
+
+    assert cost_with_whitespace == pytest.approx(len(prompt_with_whitespace) * input_cost_per_character)
+    # billing must scale with the raw string length, not the whitespace-stripped length
+    assert cost_with_whitespace > cost_without_whitespace
+
+
 def test_completion_cost_anthropic():
     """
     model_name: claude-3-haiku-20240307