Hi all,
so I have immediately tried gpt-5 family of models, and I have to say… I am disappointed 
They are about 2-3 times slower than gpt-4.1 for me. I have quite a long prompt, and a simple question “Tell me something about penguins” takes 2.5s on gpt-4.1 and 5s on gpt-5-nano (up to 8s on gpt-5).
The gpt-5-chat does not even work for me, but I assume it will be same as gpt-5.
What are your observations?
I’ve been playing with Horizon Beta, which I believe is supposed to be a smaller variant. I agree, disappointing. I haven’t let it loose on the house yet, but its response to silly questions is nearly always to ask for clarification. Speed is pretty good, though.
How did you get it working?
Through the default OpenAI integration?
I only get Response or Intent errors from it.
I personally still don’t have access. Api or in client.
Im taking the time to build up the CUDA core for Fridays local stack and putting gpt-oss:20b on it.
Yes, I use the default integration. Am on 2025.8. I just changed my model from gpt-4.1 to gpt-5 and it works out of the box. Btw, according to the official release announcement
While GPT‑5 in ChatGPT is a system of reasoning, non-reasoning, and router models, GPT‑5 in the API platform is the reasoning model that powers maximum performance in ChatGPT. Notably, GPT‑5 with minimal reasoning is a different model than the non-reasoning model in ChatGPT, and is better tuned for developers. The non-reasoning model used in ChatGPT is available as
gpt-5-chat-latest.
I think the OpenAI integration needs an update which allows us to set the reasoning detail (and set to minimum for faster responses).
The gpt-5-chat-latest model gives me always Error talking to OpenAI not sure why. Other gpt-5 models work.
Thats how they did it when 4.1 came out.
I see the error for gpt-5-chat:
2025-08-07 22:17:02.597 ERROR (MainThread) [homeassistant.components.openai_conversation] Error talking to OpenAI: Error code: 400 - {'error': {'message': 'Functions are not supported with gpt-5-chat-latest.', 'type': 'invalid_request_error', 'param': 'tools', 'code': None}}
Not sure if I can do anything about this… Seems this model is not usable for HA?
hmm I dont believe 4.1 has reasoning capabilities. Only the oX models (and now gpt5)
Correct that model is not a tool user you will want 5.1 mini or gpt5.1-nano (probably nano)
And 4.1 totally is a light reasoner that can do recursive calls. I haven’t found stiff 4.1 can’t do in HA with good grounding. It’s just not really chatty about it. And I’m holding my opinion until I’ve seen it with good grounding. Because I can make 4 1 do some damn amazing stuff.
Ah, I also don’t see it in ChatGPT with the same account.
So it’s maybe not rolled out for all at once?
How did you check with the API, do you also only get an “Error talking to OpenAI”?
The debug view of assist doesn’t show anything helpful about the error.
You spent 20 minutes trying 5, along with everyone else in the world doing the same thing, and posted a conclusion?
5 will be just as fast as the other frontier model, with speed primary determined by model size choice and reasoning setting. That is returning a typical HA related query in a second or two. Many people’s HA queries are spending more time on STT/TTS than the AI model.
The mid model is usually the correct one for HA, in my experience. Gemini 2.5 flash is worth trying.
I’ve spent a couple of hours with Horizon Beta. I did not ask it general questions that require knowledge (sorry, no penguins). What I did do was use the 100% standard HA prompt, supply it with over 200 selected entities and asked al sorts of things about that. And that was where I believe it shines.
Things that stood out: I could ask if our house was safe and I got a good answer. It mentioned the dishwasher door was open but said that this was just for info, it because was not a problem. It knew only doors in the home were relevant, unless I talked about the car. Then it knew about the car doors too. It grasped the difference about indoor and outdoor: it noted having lights on outside helped for safety too.
When I asked about air quality, I got the right info too. It knows what CO2 levels are safe. When I asked about allergies, it used data from pollen sensors (as well as provide some general info).
I could ask numbers of doors and average temperatures. It did not matter that not all my door and window sensors are not the typical binary sensors but sensors (they distinguish smal openings or wide open). I could ask for the hottest room. It knew a room had multiple temperature sensors and provided a range. When I asked it about open blinds, and told it to consider anything below 15 percent open as closed, it did that. It knew 15% open is 85% closed.
I have sensors with day counts to a birthday. It not only understood it enough to say who has a birthday next month - it also knew to add the days to today to reconstruct the actual date. It knew it was unable to tell age without the year of birth, so it asked for that.
Mind you: I did no crafty prompting, did not supply any additional tools. This is what a regular user would get without particular AI knowledge.
So my guess is it is geared toward tool use, math and reasoning more than it is about ready knowledge. Which is fine for HA. If the penguin question was asked through HA, I would not be surprised if it took more time because HA did not have penguin sensors and it preferred to use the tools.
In those couple of hours, it never hallucinated or gave wrong answers. I did hit a rate limit. Then I stopped.
As for speed: With only very few entities answers were almost instant. With lots of entities I got answers in 1 or 2 seconds, the wait was not in any way a hindrance. but I have no clue what hardware or model size was behind it all.
It’s also working for me now.
But only if I remove the part of the prompt about my tools and stop exposing the scripts.
Is tool use working for anyone in HA with 5/mini/nano or might it need an update of the OpenAI integration?
I thought would give gpt5 a try but ran into some problems.
I setup a subentry to use gpt-5-mini but when I try a simple question, it errors and the logs says:
[homeassistant.components.openai_conversation] Error talking to OpenAI: Error code: 400 - {'error': {'message': 'Your organization must be verified to stream this model. Please go to: https://platform.openai.com/settings/organization/general and click on Verify Organization. If you just verified, it can take up to 15 minutes for access to propagate.'
OK I go to that recommended webpage and the page has a button which says:
Verifications
Verify your organization to access protected models
and I found an article explaining such:
https://help.openai.com/en/articles/10910291-api-organization-verification
So just wanted to ask if I am missing something, or is the verification the norm in this case (as I didn’t have to do this for gpt4)?
For now I suggest using OpenAI Conversation which should load when adding the API key to the main integration.
Sorry, not following … I have been using the HA native OpenAI Integration for some time and it is already setup for a Conversation to use gpt4 mini.
What is not working is the newly created (sub)entry gpt-5-mini at the bottom.
From what I have been reading, GPT5 models in streaming mode require organization verification but non-streaming mode does not.
Just glancing at the code, it seems HA is setup with streaming mode set to true.
model_args = {
"model": options.get(CONF_CHAT_MODEL, RECOMMENDED_CHAT_MODEL),
"input": [],
"max_output_tokens": options.get(CONF_MAX_TOKENS, RECOMMENDED_MAX_TOKENS),
"top_p": options.get(CONF_TOP_P, RECOMMENDED_TOP_P),
"temperature": options.get(CONF_TEMPERATURE, RECOMMENDED_TEMPERATURE),
"user": chat_log.conversation_id,
"store": False,
"stream": True,
}
I wanted to try GPT-5 Nano to see if it performs better than 4.1-mini, which I’m currently using productively with my Voice Assistant.
However, I haven’t been able to access any GPT-5 models—whether via native integration, Extended OpenAI Conversation, or the OpenRouter integration. I keep running into the same “Intent errors” described earlier in the thread, while 4.1 works perfectly with the same integrations and APIs/prompts.
My organization is verified directly with OpenAI, yet I still see this error in the logs. I also tested the same API in OpenWebUI with Stream Mode enabled, and it worked fine there:
[homeassistant.components.openai_conversation] Error talking to OpenAI: Error code: 400 – {‘error’: {‘message’: ‘Your organization must be verified to stream this model. Please go to: https://platform.openai.com/settings/organization/general and click on Verify Organization. If you just verified, it can take up to 15 minutes for access to propagate.’}}
According to OpenAI’s release notes, GPT-5 introduces new API features, such as the reasoning.effort parameter with a new minimal setting (in addition to low, medium, and high). I suspect the integrations will need an update before GPT-5 works properly. For function calls/tools, I’d expect the minimal setting to be particularly interesting for reducing latency.
Source: https://platform.openai.com/docs/guides/latest-model
There is a merged PR about GTP5 in the upcoming 2025.8.1 release:
gtp-5-nano is working for me now after the update. But mini still returns the same error.
Maybe someone else can check that too …
I’m using mini successfully