diff --git a/ms_agent/llm/anthropic_llm.py b/ms_agent/llm/anthropic_llm.py
index 00e518ea3..f1a3c23ba 100644
--- a/ms_agent/llm/anthropic_llm.py
+++ b/ms_agent/llm/anthropic_llm.py
@@ -1,7 +1,6 @@
 import inspect
 from typing import Any, Dict, Generator, Iterator, List, Optional, Union
 
-import json5
 from ms_agent.llm import LLM
 from ms_agent.llm.utils import Message, Tool, ToolCall
 from ms_agent.utils import assert_package_exist, retry
@@ -109,10 +108,20 @@ def _call_llm(self,
         if formatted_messages[0]['role'] == 'system':
             system = formatted_messages[0]['content']
             formatted_messages = formatted_messages[1:]
+
+        max_tokens = kwargs.pop('max_tokens', 16000)
+        extra_body = kwargs.get('extra_body', {})
+        enable_thinking = extra_body.get('enable_thinking', False)
+        thinking_budget = extra_body.get('thinking_budget', max_tokens)
+
         params = {
             'model': self.model,
             'messages': formatted_messages,
-            'max_tokens': kwargs.pop('max_tokens', 1024),
+            'max_tokens': max_tokens,
+            'thinking': {
+                'type': 'enabled' if enable_thinking else 'disabled',
+                'budget_tokens': thinking_budget
+            }
         }
 
         if system:
@@ -163,6 +172,8 @@ def _stream_format_output_message(self,
         )
         tool_call_id_map = {}  # index -> tool_call_id (用于去重 yield)
         with stream_manager as stream:
+            full_content = ''
+            full_thinking = ''
             for event in stream:
                 event_type = getattr(event, 'type')
                 if event_type == 'message_start':
@@ -170,8 +181,13 @@ def _stream_format_output_message(self,
                     current_message.id = msg.id
                     tool_call_id_map = {}
                     yield current_message
-                elif event_type == 'text':
-                    current_message.content = event.snapshot
+                elif event_type == 'content_block_delta':
+                    if event.delta.type == 'thinking_delta':
+                        full_thinking += event.delta.thinking
+                        current_message.reasoning_content = full_thinking
+                    elif event.delta.type == 'text_delta':
+                        full_content += event.delta.text
+                        current_message.content = full_content
                     yield current_message
                 elif event_type == 'message_stop':
                     final_msg = getattr(event, 'message')
diff --git a/tests/llm/test_anthropic.py b/tests/llm/test_anthropic.py
index ba434f4b9..cec6a381d 100644
--- a/tests/llm/test_anthropic.py
+++ b/tests/llm/test_anthropic.py
@@ -9,7 +9,7 @@
 
 from modelscope.utils.test_utils import test_level
 
-API_CALL_MAX_TOKEN = 50
+API_CALL_MAX_TOKEN = 500
 
 
 class OpenaiLLM(unittest.TestCase):
@@ -124,34 +124,23 @@ def test_tool_no_stream(self):
         print(res)
         assert (len(res.tool_calls))
 
-    @unittest.skipUnless(test_level() >= 0, 'skip test in current test level')
-    def test_agent_multi_round(self):
-        import asyncio
-
-        async def main():
-            agent = LLMAgent(config=self.conf, mcp_config=self.mcp_config)
-            if hasattr(agent.config, 'callbacks'):
-                agent.config.callbacks.remove('input_callback')  # noqa
-            res = await agent.run('访问www.baidu.com')
-            print(res)
-            assert ('robots.txt' in res[-1].content)
-
-        asyncio.run(main())
-
     @unittest.skipUnless(test_level() >= 1, 'skip test in current test level')
-    def test_stream_agent_multi_round(self):
+    def test_stream_agent_multi_round_with_thinking(self):
         import asyncio
         from copy import deepcopy
 
         async def main():
             conf2 = deepcopy(self.conf)
-            conf2.generation_config.stream = True
+            conf2.llm.model = 'Qwen/Qwen3-235B-A22B'
+            conf2.generation_config.extra_body.enable_thinking = True
             agent = LLMAgent(config=conf2, mcp_config=self.mcp_config)
             if hasattr(agent.config, 'callbacks'):
                 agent.config.callbacks.remove('input_callback')  # noqa
-            res = await agent.run('访问www.baidu.com')
-            print('res:', res)
-            assert ('robots.txt' in res[-1].content)
+            res = await agent.run('访问www.baidu.com', stream=True)
+            async for chunk in res:
+                print('res: ', chunk)
+            assert ('robots.txt' in chunk[-1].content)
+            assert (chunk[-1].reasoning_content)
 
         asyncio.run(main())