Claude 4.5 Sonnet fails with 'temperature and top_p cannot both be specified' error

[nicotem]

Since Opus 4.1 Anthropic has been enforcing its recommendation that temperature and top_p should not be both specified.
I can make Sonnet 3.7 work ok but Claude 4.* breaks no matter what I try.

I am using QC 3.7. I have been unable to configure config.ini in a way that will make Claude 4.* work, because the QC will default overwrite config.ini values for temperature and top_p that it does not like, and sends both, triggering the event.

Here below a more detailed breakdown...

Steps to Reproduce:

Configure Qualcoder with Anthropic API key in config.ini
Set up Claude 4.5 Sonnet model:

ini[ai_model_Anthropic Claude]
large_model = claude-sonnet-4-5-20250514
fast_model = claude-haiku-4-5-20251001
api_base = https://api.anthropic.com/v1/
api_key = [your-key]

Ensure config.ini has both temperature and top_p set in [DEFAULT]:

iniai_temperature = 1.0
ai_top_p = 1.0

4. Attempt to use any AI feature (AI coding, AI chat, etc.)

**Expected Behavior:**
AI features should work with Claude 4.5 models, just as they do with Claude 3.7.

**Actual Behavior:**
The following error occurs:

File "qualcoder/ai_async_worker.py", line 108, in run
File "qualcoder/ai_llm.py", line 589, in _ai_async_stream
File "langchain_core/language_models/chat_models.py", line 505, in stream
File "langchain_openai/chat_models/base.py", line 1044, in _stream
File "openai/_utils/_utils.py", line 287, in wrapper
File "openai/resources/chat/completions/completions.py", line 1087, in create
File "openai/_base_client.py", line 1256, in post
File "openai/_base_client.py", line 1044, in request

{'code': 'invalid_request_error', 'message': 'temperature and top_p cannot both be specified for this model. Please use only one.', 'type': 'invalid_request_error', 'param': None}

Error communicating with Anthropic Claude
BadRequestError: Error code: 400 - {'error': {'code': 'invalid_request_error', 'message': 'temperature and top_p cannot both be specified for this model. Please use only one.', 'type': 'invalid_request_error', 'param': None}}
Root Cause Analysis:
Anthropic introduced a breaking API change with Claude 4.1 Opus (August 2025) and Claude 4.5 Sonnet (September 2025) that prohibits specifying both temperature and top_p parameters simultaneously. Previous Claude models (3.7, 3.5, etc.) allowed both parameters, though Anthropic's documentation always recommended using only one.
This is a widespread issue affecting many applications: see n8n Issue #18304, LiteLLM Issue #15097, and LangChain.js Issue #9205.
Anthropic's official guidance now states: "You usually only need to use temperature" and designates top_p as "recommended for advanced use cases only." See: https://docs.anthropic.com/en/api/messages
Current Workaround:
Users can temporarily switch back to Claude 3.7 Sonnet which accepts both parameters:
inilarge_model = claude-3-7-sonnet-20250219
fast_model = claude-3-7-sonnet-20250219
Attempted User-Side Fixes (All Failed):

Commenting out or deleting ai_top_p line: Qualcoder's code regenerates the line in config.ini
Setting ai_top_p = None: Causes ValueError: could not convert string to float: 'None'
Leaving ai_top_p empty: Qualcoder writes ai_top_p = . which causes ValueError: could not convert string to float: '.'

The configuration file approach doesn't work because Qualcoder's code actively maintains these parameters.
Suggested Fix:
The issue is likely in qualcoder/ai_llm.py where the langchain ChatAnthropic client is initialized. The code probably looks something like:
pythonllm = ChatAnthropic(
model=model_name,
temperature=float(config['ai_temperature']),
top_p=float(config['ai_top_p']), # This causes the problem
...
)
Recommended solution: Check the model version and conditionally omit top_p for Claude 4.x models:
pythonllm_params = {
'model': model_name,
'temperature': float(config['ai_temperature']),
'api_key': api_key,
'base_url': base_url,
# ... other required params
}

Only add top_p for Claude models that support it ?
Claude 4.x models (4.1 Opus, 4.5 Sonnet, 4.5 Haiku) reject dual parameters
model_lower = model_name.lower()
if not any(x in model_lower for x in ['claude-4', 'claude-sonnet-4', 'claude-haiku-4', 'claude-opus-4']):
if config.get('ai_top_p') and config['ai_top_p'] not in [None, '', '.']:
try:
llm_params['top_p'] = float(config['ai_top_p'])
except ValueError:
pass # Skip invalid top_p values

llm = ChatAnthropic(**llm_params)
Alternative simpler fix: Just remove top_p entirely from the Anthropic integration. According to Anthropic's documentation, temperature alone is sufficient for most use cases, and top_p is only for advanced scenarios. Since Qualcoder's AI features work well with temperature alone, omitting top_p wouldn't impact functionality:
pythonllm = ChatAnthropic(
model=model_name,
temperature=float(config['ai_temperature']),
# Remove top_p parameter entirely for Anthropic models
...
)
Additional Context:

This only affects Anthropic Claude models; OpenAI and other providers continue to accept both parameters
The latest langchain-anthropic library (v0.3+) has updated defaults for Claude 4.x models where top_p defaults to null instead of -1, but Qualcoder may need to update dependencies or handle this explicitly
Anthropic's reasoning for the restriction is that combining temperature and top_p creates unpredictable sampling behavior

Impact:
This prevents users from accessing Claude 4.5 Sonnet, which is Anthropic's most capable model as of October 2025. It significantly impacts the quality of AI-assisted qualitative coding features.
Thank you for considering this fix! The Qualcoder AI integration is fantastic, and Claude 4.5 support would make it even better for qualitative researchers.

[kaixxx]

I've discovered the same problem today and already fixed it (by omitting top_p altogether for all models by Anthropic): 0c0f347

Unfortunately, you must run QualCoder from source to take advantage of it - or wait a couple of weeks until we release a new compiled version.

[kaixxx]

In the above quoted commit, I also define Claude Opus 4.1 as the new default model for Anthropic. Reading your comments, it might be better to choose Sonnet 4.5. What do you think?

[nicotem]

Thank you for the quick reply! Awesome to hear the solution is already available. I think I will switch to running from source for the time being to take advantage of it. I was hoping to get this fixed asap to run a few coding tasks through it.

[nicotem]

I think my current default preference for Claude is going to be Sonnet 4.5, because of speed and especially, cost. It might be better at operationalising instructions and following langchain, but I am conjecturing.
But along the same line I was also adding a Haiku-only entry too to drive the cost even lower. I am not sure how much I would use Opus, though the volume of work I do is small and I am not sure I would even notice the difference in terms of cost...

[nicotem]

My understanding is that Opus is going to be better at creative thinking, while Sonnet might be more rigorous in technical implementation. Then what kind of profile is qualitative analysis, I guess it falls somewhere in the middle depending on the nuance of the task at hand