Seeing as how:
``
“message” : “Unsupported parameter: ‘temperature’ is not supported with this model.”
```
is now a thing with GPT-5 models (as they’re reasoning models) - how does that work when it might get handed off to a “non-thinking” model (as is done in ChatGPT)? Or is this just not a thing anymore unless we’re specifically reaching out to 4.1
Yes, just ran into this myself. My guess is “this is just not a thing anymore,” but perhaps we’ll see.
Is it because of maintaining a stable reasoning? Will they add the ability of temperature being a parameter in the future?
Will they add the ability of temperature being a parameter in the future?
Don’t count on it. Temperature and top_p parms are not needed for GPT-5 models.
This is a breaking change and absolutely unexpected, I really hope the add it as a significant number of real-world applications rely on more deterministic responses.
I am facing the same issue! But it works with the default value of 1.0 but no option to vary it.
Does anyone else experience the same?
You experience the same behavior with other API reasoning models and attempts at top_p or temperature. You can send the defaults=1 successfully. Divergence from that will fail validation.
The playground is what is exposing an error an API programmer would rarely make with this understanding. You can easily switch model there, and be shown a call that will fail. I expect this would be tweaked up as things settle in.
One has to assume some internal task-based self-adjustment replacing your desired control.
max_tokens is not remapped either, because max_completion_tokens reflects a budget and not just what you receive.
I’m using GPT-4.1 for RP/world-building, and the “temperature” parameter has been very important for the quality of output. It even took a while to figure out the right “temperature” for each project. It’s such a big parameter that I find it odd that there’s no mention of it in the docs related to GPT-5 or the migration to it.
This is extremely annoying but what I found even more annoying is how OpenAI does not have a proper way for developers to give feedback about API issues like this one.
This was a bit of a surprise. I was unable. to find the fact that temp and top-p are not available for the GPT-5 series of models in the docs, which threw the update of my code. Working in healthcare setting, this is somewhat of a concern as locking down variability of responses is a valuable control for clinical risk.
After some digging and testing, I realized that GPT-5 does not support “temperature” at all, and it performs awfully at storytelling. However, GPT-5-chat-latest does support “temperature” and does better at storytelling but is not as intelligent as GPT-4.1 (GPT-5-chat-latest says a few dumb things that I never see GPT-4.1 say).
I’m sticking to GPT-4.1 for my storytelling app and hope it won’t be deprecated anytime soon!
GPT-5 has been a disappointment so far and seems to be for (vibe) coding mostly. Ugh!
If you know about storytelling, prime it with a storytelling model based on literature in communities of writing.
I was facing issue with GPT-5 until I see your post
I notice that GPT-5 not work if you add temperature
but GPT-5 mini it working even with temperature
I don’t know why only with the orignal model not working but mini and nano working fine
Both temperature and presence_penalty are apparently no longer supported. Why would they remove these parameters? They are vitally important for achieving certain types of content.
Incorrect, gpt-5-nano does not work with temperature or presence_penalty or frequency_penalty from what I can tell.
It’s not a replacement for gpt-4.1 - it is a reasoning model and it has a shorter input context.
Long live 4.1! 
Imma just leave this here…
from openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model="gpt-5-chat-latest",
store=False,
frequency_penalty=0.05,
presence_penalty=0.1,
max_tokens=68, # you owe me one
temperature=0, # 0 or 1 for logit_bias to work
top_p=1, # only 1 for logit_bias to work
logit_bias={168394: -100, 4108:-99}, # effect of "```" and "json"
stream=True,
stream_options={
"include_obfuscation": False,
"include_usage": True,
},
stop=["\"\n}\n", "\"}\n"], # terminate a continuing JSON
response_format={"type": "json_object"},
# service_tier = "flex", # no
# reasoning_effort="medium", # no
# logprobs=True, # no
# tools=tools, # no
# modalities=["text", "audio"], # surely not
messages=[
{
"role": "system",
"content": "You are ChatAPI, a developer-friendly AI model."
},
{
"role": "user",
"content": [
{
"type": "text",
"text": "Send a JSON, with key 'chat_introduction'"
}
]
},
],
)
assistant_content = ""
for chunk in response:
if chunk.choices and chunk.choices[0].delta.content \
and not chunk.usage:
assistant_content += chunk.choices[0].delta.content
print(chunk.choices[0].delta.content, end="")
elif chunk.usage:
print("\n" + str(chunk.usage.model_dump()))
ChatGPT-5. No reasoning. Likes markdown.
From my testing of temperature:
GPT-5: not supported 
GPT-5-chat: supported 
GPT-5-mini: not supported
GPT-5-nano: not supported
You are correct!
ChatGPT-5 will simply route your request to the right model, depending on the type of request.
One the API however, you’ll need to select the suitable model manually:
gpt-5 is similar to o4-mini. (Reasoning, no temperature)
gpt-5-chat-latest is similar to gpt-4o. (Everyday chatting, with temperature)
gpt-4.1 is similar to gpt-4.5-preview. (Writing and storytelling, with temperature)
Also requires paying a premium price for something that ancient models can do, though I guess that’s sort of the point. GPT-5 was pretty clearly designed to increase OpenAI’s bottom line, and it definitely excels in that regard.
Leaving reason on default with gpt-5-nano, it costs almost 2000 tokens to generate a basic list of 6 phrases. Pretty insane. With gpt-4.1-nano, it takes 200 tokens. gpt-5 with reasoning_effort minimal, the outputs are inferior to gpt-4.1-nano.