diff --git a/litellm-rust/crates/llms/src/anthropic/messages/thinking.rs b/litellm-rust/crates/llms/src/anthropic/messages/thinking.rs
index f964c797815..e21711dc40d 100644
--- a/litellm-rust/crates/llms/src/anthropic/messages/thinking.rs
+++ b/litellm-rust/crates/llms/src/anthropic/messages/thinking.rs
@@ -250,7 +250,13 @@ fn translate_reasoning_effort(
         });
     }
     let Some(budget) = fit_budget_to_max_tokens(budget, request.params.max_tokens) else {
-        return Ok(request);
+        if request.params.thinking.is_some() {
+            return Ok(request);
+        }
+        return Ok(MessagesRequest {
+            params: without_thinking(request.params, context),
+            ..request
+        });
     };
     let enabled = ThinkingConfig::enabled(budget);
     Ok(MessagesRequest {
@@ -687,6 +693,20 @@ mod tests {
         serde_json::json!({"max_tokens": 8192, "reasoning_effort": "high"}),
         true
     )]
+    #[case::reasoning_effort_budget_cannot_fit(
+        haiku_4_5(),
+        serde_json::json!({"max_tokens": 1024, "reasoning_effort": "high"}),
+        false
+    )]
+    #[case::caller_thinking_kept_when_budget_cannot_fit(
+        haiku_4_5(),
+        serde_json::json!({
+            "max_tokens": 1024,
+            "reasoning_effort": "high",
+            "thinking": {"type": "enabled", "budget_tokens": 1000}
+        }),
+        true
+    )]
     #[case::effort_model_drops_adaptive(opus_4_5(), claude_code_payload("high", 8192), false)]
     #[case::adaptive_model_keeps_adaptive(sonnet_4_6(), claude_code_payload("high", 8192), true)]
     #[case::non_reasoning_model(haiku_3_5(), claude_code_payload("high", 8192), false)]
diff --git a/litellm/llms/anthropic/messages/native/transformation.py b/litellm/llms/anthropic/messages/native/transformation.py
index 6869c3d4095..51c9621b7ff 100644
--- a/litellm/llms/anthropic/messages/native/transformation.py
+++ b/litellm/llms/anthropic/messages/native/transformation.py
@@ -459,6 +459,8 @@ class AnthropicMessagesConfig(BaseAnthropicMessagesConfig):
         fitted_thinking: Final = AnthropicConfig.cap_thinking_budget_to_max_tokens(mapped_thinking, max_tokens)
         if fitted_thinking is None:
             verbose_logger.warning(DROP_UNFITTING_REASONING_EFFORT_WARNING, reasoning_effort, model, max_tokens)
+            if "thinking" not in optional_params:
+                self._drop_thinking(optional_params, model, custom_llm_provider)
             return
 
         optional_params.setdefault("thinking", fitted_thinking)
diff --git a/tests/unit/llms/anthropic/messages/native/test_anthropic_messages_effort.py b/tests/unit/llms/anthropic/messages/native/test_anthropic_messages_effort.py
index 37592a66c9f..b1eed12b4a3 100644
--- a/tests/unit/llms/anthropic/messages/native/test_anthropic_messages_effort.py
+++ b/tests/unit/llms/anthropic/messages/native/test_anthropic_messages_effort.py
@@ -347,6 +347,18 @@ CLEAR_TOOL_USES_EDIT: Final = {"type": "clear_tool_uses_20250919"}
     [
         pytest.param("claude-sonnet-4-5", {"reasoning_effort": "none"}, False, id="reasoning-effort-none"),
         pytest.param("claude-haiku-4-5", {"reasoning_effort": "high"}, True, id="reasoning-effort-high"),
+        pytest.param(
+            "claude-haiku-4-5",
+            {"max_tokens": 1024, "reasoning_effort": "high"},
+            False,
+            id="reasoning-effort-budget-cannot-fit",
+        ),
+        pytest.param(
+            "claude-haiku-4-5",
+            {"max_tokens": 1024, "reasoning_effort": "high", "thinking": {"type": "enabled", "budget_tokens": 1000}},
+            True,
+            id="caller-thinking-kept-when-budget-cannot-fit",
+        ),
         pytest.param("claude-opus-4-5", _claude_code_payload(effort="high"), False, id="effort-model-drops-adaptive"),
         pytest.param("claude-sonnet-4-6", _claude_code_payload(effort="high"), True, id="adaptive-model"),
         pytest.param("claude-3-5-haiku-latest", _claude_code_payload(), False, id="non-reasoning-model"),