fix(caching): updating gemini-1.5 minimum token requirements

litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py

@@ -326,7 +326,7 @@ class LiteLLMAnthropicMessagesAdapter:
                 target["cache_control"] = cache_control  # type: ignore[typeddict-item]
             else:
                 # Fallback for non-dict objects (shouldn't happen in practice)
-                cast(Dict[str, Any], target)["cache_control"] = cache_control
+                setattr(target, "cache_control", cache_control)
 
     def translatable_anthropic_params(self) -> List:
         """

litellm/llms/vertex_ai/context_caching/transformation.py

@@ -174,14 +174,18 @@ def get_gemini_context_caching_min_tokens(model: str) -> int:
 
     Gemini rejects a cachedContents create below a per-model floor with a 400, so
     the caller skips caching below this value. Figures from
-    https://ai.google.dev/gemini-api/docs/caching (Gemini 2.5 -> 2048, Gemini 3.x
-    -> 4096). Unknown Gemini models default to the highest known floor so a create
-    is never attempted below the real minimum.
+    https://ai.google.dev/gemini-api/docs/caching (Gemini 1.5 -> 32768, Gemini 2.5
+    -> 2048, Gemini 3.x -> 4096). Unknown Gemini models default to the highest
+    known floor so a create is never attempted below the real minimum.
     """
     model_lower = model.lower()
+    if "gemini-1.5" in model_lower or "gemini-1-5" in model_lower:
+        return 32768
     if "gemini-2.5" in model_lower or "gemini-2-5" in model_lower:
         return 2048
-    return 4096
+    if "gemini-3" in model_lower:
+        return 4096
+    return 32768
 
 
 def separate_cached_messages(

tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py

@@ -1431,6 +1431,21 @@ def test_should_add_cache_control_for_gemini_model():
         assert target.get("cache_control") == cache_control
 
 
+def test_cache_control_fallback_setattr():
+    """Verify cache_control is safely assigned to non-dict target objects using setattr."""
+    adapter = LiteLLMAnthropicMessagesAdapter()
+    cache_control = {"type": "ephemeral"}
+
+    class MockTarget:
+        pass
+
+    target = MockTarget()
+    adapter._add_cache_control_if_applicable(
+        {"cache_control": cache_control}, target, "claude-3-opus-20240229"
+    )
+    assert getattr(target, "cache_control", None) == cache_control
+
+
 def test_cache_control_preserved_in_text_content_for_gemini():
     """cache_control must survive message translation for a Gemini target."""
     anthropic_messages = [

tests/test_litellm/llms/vertex_ai/context_caching/test_context_caching_ttl.py

@@ -14,6 +14,9 @@ class TestGeminiContextCachingMinTokens:
     @pytest.mark.parametrize(
         "model, expected",
         [
+            ("gemini-1.5-pro", 32768),
+            ("gemini-1.5-flash", 32768),
+            ("vertex_ai/gemini-1.5-pro-001", 32768),
             ("gemini-2.5-flash", 2048),
             ("gemini-2.5-pro", 2048),
             ("gemini/gemini-2.5-pro", 2048),
@@ -21,7 +24,7 @@ class TestGeminiContextCachingMinTokens:
             ("gemini-3.5-flash", 4096),
             ("gemini-3.1-pro-preview", 4096),
             ("gemini/gemini-3.5-flash", 4096),
-            ("gemini-1.5-pro", 4096),
+            ("gemini-unknown-future-model", 32768),
         ],
     )
     def test_min_tokens_by_model(self, model, expected):

tests/test_litellm/llms/vertex_ai/context_caching/test_vertex_ai_context_caching.py

@@ -1407,6 +1407,7 @@ class TestContextCachingEndpoints:
             ("gemini-3.5-flash", 4096),
             ("gemini/gemini-3.5-flash", 4096),
             ("gemini-3.1-pro-preview", 4096),
+            ("gemini-1.5-pro", 32768),
             ("gemini-2.5-flash", 2048),
             ("gemini-2.5-pro", 2048),
         ],