fix(anthropic): disable extended thinking for Opus 4.1 (#10532)

Co-authored-by: openhands <openhands@all-hands.dev>

openhands/llm/llm.py

@@ -166,6 +166,20 @@ class LLM(RetryMixin, DebugMixin):
         elif 'gemini' in self.config.model.lower() and self.config.safety_settings:
             kwargs['safety_settings'] = self.config.safety_settings
 
+        # Explicitly disable Anthropic extended thinking for Opus 4.1 to avoid
+        # requiring 'thinking' content blocks. See issue #10510.
+        if 'claude-opus-4-1' in self.config.model.lower():
+            kwargs['thinking'] = {'type': 'disabled'}
+
+        # Anthropic constraint: Opus models cannot accept both temperature and top_p
+        # Prefer temperature (drop top_p) if both are specified.
+        _model_lower = self.config.model.lower()
+        # Limit to Opus 4.1 specifically to avoid changing behavior of other Anthropic models
+        if ('claude-opus-4-1' in _model_lower) and (
+            'temperature' in kwargs and 'top_p' in kwargs
+        ):
+            kwargs.pop('top_p', None)
+
         self._completion = partial(
             litellm_completion,
             model=self.config.model,

openhands/llm/model_features.py

@@ -100,7 +100,6 @@ REASONING_EFFORT_PATTERNS: list[str] = [
     'gemini-2.5-pro',
     'gpt-5',
     'gpt-5-2025-08-07',
-    'claude-opus-4-1-20250805',
     # DeepSeek reasoning family
     'deepseek-r1-0528*',
 ]

tests/unit/llm/test_llm.py

@@ -1204,7 +1204,7 @@ def test_gemini_medium_reasoning_effort_passes_through(mock_completion):
 
 
 @patch('openhands.llm.llm.litellm_completion')
-def test_opus_41_reasoning_pops_temperature_top_p(mock_completion):
+def test_opus_41_keeps_temperature_top_p(mock_completion):
     mock_completion.return_value = {
         'choices': [{'message': {'content': 'ok'}}],
     }
@@ -1217,7 +1217,8 @@ def test_opus_41_reasoning_pops_temperature_top_p(mock_completion):
     llm = LLM(config, service_id='svc')
     llm.completion(messages=[{'role': 'user', 'content': 'hi'}])
     call_kwargs = mock_completion.call_args[1]
-    assert 'temperature' not in call_kwargs
+    assert call_kwargs.get('temperature') == 0.7
+    # Anthropic rejects both temperature and top_p together on Opus; we keep temperature and drop top_p
     assert 'top_p' not in call_kwargs
 
 
@@ -1239,6 +1240,21 @@ def test_opus_4_keeps_temperature_top_p(mock_completion):
     assert call_kwargs.get('top_p') == 0.9
 
 
+@patch('openhands.llm.llm.litellm_completion')
+def test_opus_41_disables_thinking(mock_completion):
+    mock_completion.return_value = {
+        'choices': [{'message': {'content': 'ok'}}],
+    }
+    config = LLMConfig(
+        model='anthropic/claude-opus-4-1-20250805',
+        api_key='k',
+    )
+    llm = LLM(config, service_id='svc')
+    llm.completion(messages=[{'role': 'user', 'content': 'hi'}])
+    call_kwargs = mock_completion.call_args[1]
+    assert call_kwargs.get('thinking') == {'type': 'disabled'}
+
+
 @patch('openhands.llm.llm.litellm.get_model_info')
 def test_is_caching_prompt_active_anthropic_prefixed(mock_get_model_info):
     # Avoid external calls, but behavior shouldn't depend on model info

tests/unit/llm/test_model_features.py

@@ -199,7 +199,6 @@ def test_function_calling_models(model):
         'gemini-2.5-flash',
         'gemini-2.5-pro',
         'gpt-5',
-        'claude-opus-4-1-20250805',
     ],
 )
 def test_reasoning_effort_models(model):
@@ -230,7 +229,6 @@ def test_deepseek_reasoning_effort_models(model):
         'claude-3-haiku-20240307',
         'claude-3-opus-20240229',
         'claude-sonnet-4-latest',
-        'claude-opus-4-1-20250805',
     ],
 )
 def test_prompt_cache_models(model):
@@ -254,7 +252,7 @@ def test_prompt_cache_models(model):
         ('gemini-2.5-pro', True),
         ('gpt-5', True),
         ('gpt-5-2025-08-07', True),
-        ('claude-opus-4-1-20250805', True),
+        ('claude-opus-4-1-20250805', False),
         # DeepSeek
         ('deepseek/DeepSeek-R1-0528:671b-Q4_K_XL', True),
         ('DeepSeek-R1-0528', True),