☠ GPT-5 update Documentation Errors

Documentation
1

:pushpin:

The reasons I am highlighting these only a few days after launch:

  • Should have been fixed/updated prior to launch (imagine someone wanting to switch over from the twitter bot to openai; it would be that much more annoying without proper docs)
  • Every time I come back to the openai docs, they are not fixed for months to match new models, only adding to time spent on deconfusing the code
  • these are the major parameter changes which make prior functions incompatible with new models
  1. Reasoning effort on Responses API:

reasoning obj
reasoning obj1190×755 107 KB

  1. Verbosity on Responses API
  • Currently supported values are low, medium, and high. ← wrong; none of the models support this, except for GPT-5 family models; o3, o4, etc only support “medium”
  • Verbosity being listed as it is now in the docs suggests it is its own parameter. It is a subparameter of text. This can also be seen in the sample Response object they provide on the left side; no it cannot, because verbosity does not exist there. 
    • verbosity = "low" # gives error

    • text          = {"verbosity": "low"} # works

verbosity obj
verbosity obj1904×1209 283 KB

  1. Temperature
  • NOT supported in the o3, o4 or new family of GPT models; it will give error unless ignored or set to 1

temperature obj
temperature obj1880×724 138 KB

  1. All of the above 1-3, but for the Chat Completions API documentation

  2. stop parameter in Chat Completions API

  • does not work for GPT 5 family (and does not even work in the Responses API at all)

stop obj
stop obj1222×574 87.1 KB

2

Documentation → models → gpt-5 and variants:

Says it supports predictions. NOT
Says it supports fine tuning. NOT
Says it supports assistants. NOT

The concerns above cannot answer about all models. For example, shall “temperature” or “logit_bias” or “input_image” in the API reference refer to every model where it is refused?

There is an immense parameter logic grid needed if you have a model selector. humanHours * everyDeveloper to re-create what OpenAI leaves centrally undocumented.

A bigger and major concern is that OpenAI just completely obliterated Chat Completions from the documentation. There is no more selector in almost every category.

4

The model gpt-5-chat-latest has more differences (besides its model page) : support temperature unlike other gpt-5 series, does not support verbosity when it should (they said that all previous models have been using verbosity medium).