diff --git a/openhands/llm/llm.py b/openhands/llm/llm.py index f97669309c..b94ed3bc2b 100644 --- a/openhands/llm/llm.py +++ b/openhands/llm/llm.py @@ -188,13 +188,13 @@ class LLM(RetryMixin, DebugMixin): 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 + # Anthropic constraint: Opus 4.1 and Sonnet 4 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 - ): + # Apply to Opus 4.1 and Sonnet 4 models to avoid API errors + if ( + ('claude-opus-4-1' in _model_lower) or ('claude-sonnet-4' in _model_lower) + ) and ('temperature' in kwargs and 'top_p' in kwargs): kwargs.pop('top_p', None) # Add completion_kwargs if present diff --git a/tests/unit/llm/test_llm.py b/tests/unit/llm/test_llm.py index a6fc7c6726..dfdb4e05b4 100644 --- a/tests/unit/llm/test_llm.py +++ b/tests/unit/llm/test_llm.py @@ -1255,6 +1255,25 @@ def test_opus_41_keeps_temperature_top_p(mock_completion): assert 'top_p' not in call_kwargs +@patch('openhands.llm.llm.litellm_completion') +def test_sonnet_4_keeps_temperature_drops_top_p(mock_completion): + mock_completion.return_value = { + 'choices': [{'message': {'content': 'ok'}}], + } + config = LLMConfig( + model='anthropic/claude-sonnet-4-20250514', + api_key='k', + temperature=0.7, + top_p=0.9, + ) + llm = LLM(config, service_id='svc') + llm.completion(messages=[{'role': 'user', 'content': 'hi'}]) + call_kwargs = mock_completion.call_args[1] + assert call_kwargs.get('temperature') == 0.7 + # Anthropic rejects both temperature and top_p together on Sonnet 4; we keep temperature and drop top_p + assert 'top_p' not in call_kwargs + + @patch('openhands.llm.llm.litellm_completion') def test_opus_4_keeps_temperature_top_p(mock_completion): mock_completion.return_value = {