diff --git a/src/api/models/bedrock.py b/src/api/models/bedrock.py
index 0aef7488..7eef15fd 100644
--- a/src/api/models/bedrock.py
+++ b/src/api/models/bedrock.py
@@ -774,8 +774,14 @@ def _parse_request(self, chat_request: ChatRequest) -> dict:
         system_prompts = self._parse_system_prompts(chat_request)
 
         # Base inference parameters.
+        # Prefer max_completion_tokens (OpenAI newer field) over max_tokens (legacy).
+        effective_max_tokens = (
+            chat_request.max_completion_tokens
+            if chat_request.max_completion_tokens is not None
+            else chat_request.max_tokens
+        )
         inference_config = {
-            "maxTokens": chat_request.max_tokens,
+            "maxTokens": effective_max_tokens,
         }
 
         # Only include optional parameters when specified
@@ -818,15 +824,11 @@ def _parse_request(self, chat_request: ChatRequest) -> dict:
 
             if "anthropic.claude" in model_lower:
                 # Claude format: reasoning_config = object with budget_tokens
-                max_tokens = (
-                    chat_request.max_completion_tokens
-                    if chat_request.max_completion_tokens
-                    else chat_request.max_tokens
-                )
+                # effective_max_tokens already prefers max_completion_tokens over max_tokens
                 budget_tokens = self._calc_budget_tokens(
-                    max_tokens, chat_request.reasoning_effort
+                    effective_max_tokens, chat_request.reasoning_effort
                 )
-                inference_config["maxTokens"] = max_tokens
+                inference_config["maxTokens"] = effective_max_tokens
                 # unset topP - Not supported
                 inference_config.pop("topP", None)
 
diff --git a/src/api/schema.py b/src/api/schema.py
index d4e39db3..ce1879a3 100644
--- a/src/api/schema.py
+++ b/src/api/schema.py
@@ -3,7 +3,7 @@
 
 from pydantic import BaseModel, Field
 
-from api.setting import DEFAULT_MODEL
+from api.setting import DEFAULT_MAX_TOKENS, DEFAULT_MODEL
 
 
 class Model(BaseModel):
@@ -106,7 +106,7 @@ class ChatRequest(BaseModel):
     temperature: float | None = Field(default=None, le=2.0, ge=0.0)
     top_p: float | None = Field(default=None, le=1.0, ge=0.0)
     user: str | None = None  # Not used
-    max_tokens: int | None = 2048
+    max_tokens: int | None = DEFAULT_MAX_TOKENS
     max_completion_tokens: int | None = None
     reasoning_effort: Literal["low", "medium", "high"] | None = None
     n: int | None = 1  # Not used
diff --git a/src/api/setting.py b/src/api/setting.py
index c69780b4..dc2e77e2 100644
--- a/src/api/setting.py
+++ b/src/api/setting.py
@@ -11,6 +11,7 @@
 
 DEBUG = os.environ.get("DEBUG", "false").lower() != "false"
 AWS_REGION = os.environ.get("AWS_REGION", "us-west-2")
+DEFAULT_MAX_TOKENS = int(os.environ.get("DEFAULT_MAX_TOKENS", "2048"))
 DEFAULT_MODEL = os.environ.get("DEFAULT_MODEL", "anthropic.claude-3-sonnet-20240229-v1:0")
 DEFAULT_EMBEDDING_MODEL = os.environ.get("DEFAULT_EMBEDDING_MODEL", "cohere.embed-multilingual-v3")
 ENABLE_CROSS_REGION_INFERENCE = os.environ.get("ENABLE_CROSS_REGION_INFERENCE", "true").lower() != "false"