diff --git a/autogpt_platform/backend/backend/copilot/config.py b/autogpt_platform/backend/backend/copilot/config.py
index bd1acbce83..b580bda08b 100644
--- a/autogpt_platform/backend/backend/copilot/config.py
+++ b/autogpt_platform/backend/backend/copilot/config.py
@@ -152,19 +152,28 @@ class ChatConfig(BaseSettings):
         "overloaded). The SDK automatically retries with this cheaper model.",
     )
     claude_agent_max_turns: int = Field(
-        default=1000,
+        default=50,
         ge=1,
         le=10000,
         description="Maximum number of agentic turns (tool-use loops) per query. "
         "Prevents runaway tool loops from burning budget.",
     )
     claude_agent_max_budget_usd: float = Field(
-        default=100.0,
+        default=5.0,
         ge=0.01,
         le=1000.0,
         description="Maximum spend in USD per SDK query. The CLI aborts the "
         "request if this budget is exceeded.",
     )
+    claude_agent_max_thinking_tokens: int = Field(
+        default=8192,
+        ge=1024,
+        le=128000,
+        description="Maximum thinking/reasoning tokens per LLM call. "
+        "Extended thinking on Opus can generate 50k+ tokens at $75/M — "
+        "capping this is the single biggest cost lever. "
+        "8192 is sufficient for most tasks; increase for complex reasoning.",
+    )
     claude_agent_max_transient_retries: int = Field(
         default=3,
         ge=0,
diff --git a/autogpt_platform/backend/backend/copilot/sdk/service.py b/autogpt_platform/backend/backend/copilot/sdk/service.py
index 35b87cd40c..5ee6bba8ca 100644
--- a/autogpt_platform/backend/backend/copilot/sdk/service.py
+++ b/autogpt_platform/backend/backend/copilot/sdk/service.py
@@ -2238,6 +2238,10 @@ async def stream_chat_completion_sdk(
             "max_turns": config.claude_agent_max_turns,
             # max_budget_usd: per-query spend ceiling enforced by the CLI.
             "max_budget_usd": config.claude_agent_max_budget_usd,
+            # max_thinking_tokens: cap extended thinking output per LLM call.
+            # Thinking tokens are billed at output rate ($75/M for Opus) and
+            # account for ~54% of total cost.  8192 is the default.
+            "max_thinking_tokens": config.claude_agent_max_thinking_tokens,
         }
         if sdk_model:
             sdk_options_kwargs["model"] = sdk_model