max_tokens Not Supported with gpt-5-nano

[ryanoun]

[Krazal]

Hotfix created; the problem can be fixed by downloading the appropriate (64-bit [x64] or 32-bit [x86]) ZIP file, unpacking NppOpenAI.dll and overwriting C:\Program Files\Notepad++\plugins\NppOpenAI\NppOpenAI.dll.*

max_tokens can still be used in the NppOpenAI.ini config file; for gpt-5 models, max_completion_tokens will be sent automatically instead.

Until a final solution (introduction of max_completion_tokens), the above release cannot be considered suitable for Release.

Download & Digest:
NppOpenAI_x64.zip: sha256:76444a79cb30349e78cd62074f6b84f1649cd28f7fe07a479daebb1f4f0f3ea8
NppOpenAI_x86.zip: sha256:fca1a514eadf4ccf970c96b1c8dfc4c7f00b7810dab9d6fc62f09b6bd8214933

* Disclaimer: The attached ZIP files may contain experimental developments. Be sure to make a backup of the previous DLL!

[ryanoun]

I'm having issues with getting a response, I think it might be due to 'top_p' getting sent even though I removed it from the config file

=== REQUEST BODY ===
{"messages":[{"content":"whats the population of the world?\n\n\n\n","role":"user"}],"model":"gpt-5-nano","top_p":0.800000011920929,"stream":true}

=== RESPONSE ===
Status: 400
Cache-Control: no-cache
Connection: keep-alive
Access-Control-Allow-Origin: *
Access-Control-Allow-Methods: GET, POST, PUT, PATCH, DELETE, OPTIONS
Access-Control-Allow-Headers: *
Content-Type: text/event-stream

When I run it in curl:

curl https://api.openai.com/v1/chat/completions
-H "Authorization: Bearer KEY"
-H "Content-Type: application/json"
-d '{
"model": "gpt-5-nano",
"messages": [
{ "role": "user", "content": "whats the population of the world?" }
],
"top_p": 0.800000011920929
}'

{
"error": {
"message": "Unsupported parameter: 'top_p' is not supported with this model.",
"type": "invalid_request_error",
"param": "top_p",
"code": "unsupported_parameter"
}

[Krazal]

For o1 / reasoning models, it is worth setting the value of top_p to 1, 1.) as stated/recommended in the OpenAI Developer Community, 2.)* and then it will NOT be sent based on the related source code.

In the future, an is_resoning_model (or similar) flag may be introduced (instead of max_completion_tokens) to ensure that the request values are either not sent or sent correctly to the related AI API.

In addition, a UI interface is planned to make configuration easier (there are also plans to switch from the current INI config to a more flexible, e.g., JSON-based configuration file).

I apologize for any inconvenience and thank you for your feedback!

* A similar problem may have existed in the case of max_tokens vs. max_completion_tokens.

[jmagnone]

I am unable to run model=gpt-5-nano because of this error. Any tentative date for a release that fixes the max_tokens and top_p Not Supported issues?

[Krazal]

Hi,

this is currently In progress on Kanban, but I moved the task to the top of the list.

Unfortunately, I've not been able to make any significant improvements/developments to the plugin in recent months, but I'm confident that this will change in the near future!