Top 10 ChatGPT‑5 User Complaints
Key Points
- The rollout of ChatGPT‑5 sparked intense backlash, not just because of the infamous “chartgate” mistake but because it abruptly terminated users’ long‑standing AI workflows and relationships built on earlier versions.
- OpenAI replaced multiple specialized models with a single “GPT‑5” that actually contains ten new sub‑models behind a router, aiming to satisfy diverse needs (speed, empathy, depth, web search) while managing GPU load.
- The router’s default to the faster, less‑reasoning sub‑model has left many users frustrated, prompting questions about when to use each variant and how to customize the experience.
- Despite widespread criticism, Sam Altman affirmed that this composite model is the permanent default for hundreds of millions of users, and the speaker outlines the top ten user complaints with practical fixes for adapting to the new system.
Sections
- Backlash Over ChatGPT 5 Rollout - The speaker details the angry user response to OpenAI abruptly replacing established AI workflows with a single GPT‑5 model that actually hides ten distinct variants, disrupting long‑term personal and professional AI relationships.
- Custom Instructions and Model Transparency - The speaker advises using explicit prompts and custom instructions to steer ChatGPT's behavior, highlights the inconsistency between chat and API model selection, and notes that true model control requires the API or Pro tier features.
- Long Context Illusion Explained - The speaker cautions that bigger token windows don’t ensure perfect recall and that traditional prompting tactics—anchoring, reiterating, and rhythmic reminders—remain crucial for handling long‑context inputs.
- Choosing ChatGPT Personality Settings - The speaker explains how to select ChatGPT’s default mode or customize its personality via the settings menu—offering options like empathetic “Listener” or “Robot”—to address complaints about reasoning depth and lack of empathy.
- Common LLM Pitfalls & Fixes - The speaker reviews four major issues—router misrouting, chat vs. API model selection, retired‑model drift, and long‑context misconceptions—and outlines practical remedies such as “think hard” prompts, custom instructions, explicit model selection, prompt versioning, and disciplined long‑context techniques.
- One Model, Many Challenges - The speaker explains the shift to a single, dominant AI model, emphasizing inevitable rollout issues and the lasting importance of prompting, model literacy, and workflow adaptation.
Full Transcript
# Top 10 ChatGPT‑5 User Complaints **Source:** [https://www.youtube.com/watch?v=Gqnf5f1ITyo](https://www.youtube.com/watch?v=Gqnf5f1ITyo) **Duration:** 00:21:59 ## Summary - The rollout of ChatGPT‑5 sparked intense backlash, not just because of the infamous “chartgate” mistake but because it abruptly terminated users’ long‑standing AI workflows and relationships built on earlier versions. - OpenAI replaced multiple specialized models with a single “GPT‑5” that actually contains ten new sub‑models behind a router, aiming to satisfy diverse needs (speed, empathy, depth, web search) while managing GPU load. - The router’s default to the faster, less‑reasoning sub‑model has left many users frustrated, prompting questions about when to use each variant and how to customize the experience. - Despite widespread criticism, Sam Altman affirmed that this composite model is the permanent default for hundreds of millions of users, and the speaker outlines the top ten user complaints with practical fixes for adapting to the new system. ## Sections - [00:00:00](https://www.youtube.com/watch?v=Gqnf5f1ITyo&t=0s) **Backlash Over ChatGPT 5 Rollout** - The speaker details the angry user response to OpenAI abruptly replacing established AI workflows with a single GPT‑5 model that actually hides ten distinct variants, disrupting long‑term personal and professional AI relationships. - [00:03:36](https://www.youtube.com/watch?v=Gqnf5f1ITyo&t=216s) **Custom Instructions and Model Transparency** - The speaker advises using explicit prompts and custom instructions to steer ChatGPT's behavior, highlights the inconsistency between chat and API model selection, and notes that true model control requires the API or Pro tier features. - [00:07:06](https://www.youtube.com/watch?v=Gqnf5f1ITyo&t=426s) **Long Context Illusion Explained** - The speaker cautions that bigger token windows don’t ensure perfect recall and that traditional prompting tactics—anchoring, reiterating, and rhythmic reminders—remain crucial for handling long‑context inputs. - [00:11:13](https://www.youtube.com/watch?v=Gqnf5f1ITyo&t=673s) **Choosing ChatGPT Personality Settings** - The speaker explains how to select ChatGPT’s default mode or customize its personality via the settings menu—offering options like empathetic “Listener” or “Robot”—to address complaints about reasoning depth and lack of empathy. - [00:16:00](https://www.youtube.com/watch?v=Gqnf5f1ITyo&t=960s) **Common LLM Pitfalls & Fixes** - The speaker reviews four major issues—router misrouting, chat vs. API model selection, retired‑model drift, and long‑context misconceptions—and outlines practical remedies such as “think hard” prompts, custom instructions, explicit model selection, prompt versioning, and disciplined long‑context techniques. - [00:19:37](https://www.youtube.com/watch?v=Gqnf5f1ITyo&t=1177s) **One Model, Many Challenges** - The speaker explains the shift to a single, dominant AI model, emphasizing inevitable rollout issues and the lasting importance of prompting, model literacy, and workflow adaptation. ## Full Transcript
So, the response to chat GPT5 has been a
little bit like watching a mob with
pitchforks come to the vampires castle.
It's been wild to see people get so
upset, so fed up with how the roll out
was handled. And I don't just mean
chartgate where famously, infamously
chat GPT5 was rolled out with completely
inaccurate charts in a live stream to
hundreds of thousands of people. That's
very fixable and OpenAI immediately
fixed it. What I mean is that they chose
to open AAI chose to end people's
long-term relationships with their AI.
And I don't just mean the sort of like
vaguely creepy this is my AI girlfriend
stuff. I mean they chose to end
workflows. They chose to end
professional engagements that people
have with thinking partners. Everything
you've built up with your AI with 40,
with 03, with 03 Pro, it all went away
within an hour or two after that video.
and instead you got a brand new AI that
was really like I've actually counted it
up 10 different GPT5 models hiding
inside the one GPT5 in a way that's
predictable right the entire world spent
a year telling open AI please stop
giving us so many models in the dropdown
but people still have really differing
needs some people want really fast
responses some people want a warm and
empathetic model some people want really
thoughtful responses some people want a
lot of inference time. Some people want
web search. Great. So, OpenAI gave us
one model that was actually 10 models
underneath with a router. And contrary
to popular belief, this is not like a
bunch of old models stitched together
with a router. These are all new models
and they're stitched in with a router.
The problem is the router is cued to
give OpenAI more room on their GPUs
because their GPUs are melting with the
kind of traffic that they get. And so
the model router defaults to the dumber
model for lack the non-reasoning model
is the polite way to put it. What do we
do with that? When do we need a
non-reasoning fast model versus a model
that's good? And how do we customize it?
This video really focuses on top 10
named user complaints and concerns and
what we can do to fix them in Chad GPT
5. I'm all about fixes, right? I'm all
about being practical. This is the new
default model for something like 700
million people whether we like it or
not. And Sam Elman on his Reddit AMA
where people came for him with
forks. He was very clear. We're not
going back. This is the model we have. I
think it's a powerful model, but I think
it needs like any model some working in
some getting to know it. It's like going
on a first date. I know that's going to
sound weird and creepy, but stick with
me, right? Andre Karpathy talked about
these as stochastic people spirits. In
this sense, you have to teach the
stochastic people spirit what you need
from it. And there are specific ways you
can do that. And so I'm going to give
you the top 10 issues that I've dug up
on the internet about chat GPT5. And I'm
going to tell you how I think they can
be addressed. And we'll go through one
at a time. Number one is router
misouting. Now, part of that on day one
was that one of their auto switch
routers was actually offline. And so if
you had day one issues but haven't had
them since, that was probably what was
going on. But if you still get shallow
responses to complex questions because
the router defaults to faster models,
you want to get to a place where you can
actually ask for thinking hard very
clearly. I would recommend two things.
One, just say think hard in the prompt.
Let's not make this over complex. And
two, go into the option to personalize
your chat GPT and make it clear in the
custom instructions what you want. as an
example, default to deep analysis unless
I say quick take and then go from there.
But essentially, you're trying to push
it and route it with the custom
instructions as much as you can. Number
two, chat versus API mismatch was also
complained. So, chat GPT uses a routing
system. API gives you direct model
access. developers get a much different
experience than the rest of us with chat
GBT5 because developers can test a
particular model in the sandbox, deploy
it and get completely different
behaviors. In this case, I think how Sam
Alman is going to address it is he's
going to start giving us more
customizability and they've already
rolled out in the last couple of hours a
ability to see what model you're getting
and what model's responding to you. And
originally that wasn't the case. So,
they're working hard to make this more
visible in the chat. That's not really
something we can fix with prompts. I
promise to be honest with you about what
you can fix and what you cannot. If you
really care about controlling exactly
which model you get every single time,
you have only a couple of options. You
can either go to the API or you can hit
the drop down and you have you don't
have 10 models of choices. You have if
you're if you're a pro user, you have
chat GPT5 pro and you have chat GBT5
thinking and you have chat GBT5. The
options degrade from there down to plus
and free users and so you have less and
less choice and have to rely more on the
prompting I gave you for router where
you're prompting think hard. This is an
issue that they are going to address
with more customization. Number three,
model drift and mismatch. So old
workflows produce different outputs
after migration to chat GPT5. That is
somewhat inevitable. I would suggest if
you've been running workflows in
production that you I I hope that you
have been keeping track of your prompts
that you have been versioning your
prompts and that when you have a new
model that is responding differently
with outputs because drift is inevitable
with new models. Any new model would
have produced drift that you then have
the space to deliberately experiment
with your prompt and adjust it to the
right model. Now, if you're running a
production pipeline, you get to select
exactly which GPT5 model you want to
use, and that gives you the flexibility
to be much more controlled in your
responses. If you're trying to run
something through the chatbot flow, and
a lot of people do, you are going to
have to do more work to customize your
prompt and more work to figure out how
to route it to the right kind of model.
And by the way, not every prompt needs a
thinking model. Sometimes you want
something quicker. I will say having
worked with this model, you sometimes
get more token output on the
non-reasoning model because it's cheaper
for them to produce those tokens. And so
if you have like a thinking model
produce an outline, you can have the
non-thinking model do a lot of work for
you in writing. Let's say you're writing
a PRD. That might be a way to do it. And
the non-thinking model, I know people
come after it. This is just a little,
you know, before we get to number four,
this a little sidebar. The non-thinking
model is remarkably smart for a
non-thinking model. And it's also
incredibly fast. And one of the things I
noticed that is true about chat GPT5
that hasn't been true about previous
models is that even if the non-thinking
model isn't right the first time, it is
so incredibly fast that you can get five
or six responses back in the time it
takes like Claude Opus 4 to do one
response. It iterates into something
that's really good in that time. And so
in in a sense people are sort of
sleeping on the value of speed there.
Okay, let's go to number four. the long
context illusion. So users have have
assumed that if they stuff the the model
with 200,000 tokens because they
advertise, right? They they advertise
they had a bigger token window that
you'll get perfect recall. It's going to
be good recall. It's going to be better
recall than we've had in the past. It
doesn't mean it's perfect. Even OpenAI's
own evaluation admits something like 89%
accuracy between 128 and 256,000 tokens.
That's that's good. It's not perfect.
There's still lost in the middle
problems. you would still be wise to use
U-shaped thinking in your prompting.
Right? So, the mitigations are not new
here. We've had challenges managing long
context windows in the past. You want to
anchor at the beginning with a strong
prompt. You want to reiterate what you
need to at the end. You can use
techniques like um rhythmic reminders
through the context window of what
you're looking for. Claude showed us
that with the system prompt. And so,
there's a lot of techniques that we
already know to manage this. And I think
people just assume they didn't have to
anymore. And as I emphasize over and
over again, these are models within a
lineage. They are getting better. But
don't assume that everything you learned
immediately breaks. Instead, assume that
a lot of the techniques you've learned
will evolve. And so in this case, it
gets a little bit easier to recall the
context. But still, those techniques
work well. Let's move on to number five.
If you ask for JSON and you just say,
"Please return JSON." For whatever
reason, chat GPT5 doesn't always do
that. Sometimes it does. Sometimes it's
invalid JSON objects. I would recommend
that you ask specifically for structured
outputs with JSON schema. And if you use
JSON a lot, I would recommend getting
into custom instructions with it and
actually specifying what you're looking
for. It's not that the system doesn't
know it. It's that for whatever reason,
each every model has flavors and tweaks.
This model in early testing has had some
issue with JSON objects. Now, that's not
for every single one. This was
specifically in some of the smaller
versions of GPT5. GPT5 Mini had this
issue. And so you might also switch to a
different model. That's going to feel
like a very coding specific tip, but we
use these models for a lot of things and
coding is one of them. Number six, tool
action and how you handle tool action
claims or calls. So the model will
sometimes pretend to have called a tool
or claim to have called a tool and done
an action it didn't perform. 03 would do
this too. An AI claims they reduce
deception significantly. Anecdotally,
that feels correct. It does do more of
what I ask it to do, but the number is
not zero. Whatever it is, I think they
claimed it it's down to 2%. It's not
zero. You need to be really clear about
requiring the model in your prompt. This
is whether you're API or chat. You need
to be clear about requiring the model to
show you a plan and then to show you the
actions completed against the plan. In
my initial notes, in my review that I
published last Friday, I talked about
the idea that this model does well with
artifacts because artifacts are a way of
proving that you can make a tool call
and come back and do something. So if
you need it to use Python, you don't
just say use Python, you say, show me
the Python greater that you made or show
me the Python query you built. So you
have to make it prove the artifact. I
think that's something that is a little
bit of a secret hack with shed GPT5
because we can't pick either the model
directly in the chat nor can we define
exactly the tool call in the chat. Those
are ways to sort of force a tool call
that get us what we want. And why does
that matter? Because this model is
designed to solve things with code. And
sometimes you get solutions with code
you wouldn't otherwise. My review on
Friday called out that it is okay at
making Gant charts just as an image. It
is really good at making Gant charts
with code and that that is a pattern
that repeats for other problems. Number
seven, thinking mode costs. Reasoning
uses a lot of tokens and a lot of time
and that is part of why it defaults to
non-reasoning. And so we have people
complaining and saying the thinking mode
takes too long given what it's giving
back. This is very much a preference. I
am actually personally very okay with
the model taking a few moments to think
before it returns because I can feel the
difference in the quality of response.
If you don't want it to think that hard,
this is actually the easiest one to
solve. Pick regular chat GPT5 or if
you're on a free uh or plus tier, it's
going to default that way anyway. And
just be happy and use that. And for a
lot of people, honestly, that is
probably good enough. By the way, the
people who complain about non-reasoning
are often complaining about either the
quality of response, and we talked about
going to thinking if you want it, or the
lack of empathy in the non-thinking. And
I have a I have a really easy fix for
you on the empathy one. Go into your
chat GPT personalization menu. You will
have a style or a mode that you can use.
And so, you can literally go in and you
can say, I'm going to actually like read
off to you all of the different options
that you can check in the settings. So
you go to settings, you go to customize
chat GPT and you can select the
personality. Personality is either
default which is quick, clever and built
to keep the conversation going which is
absolutely true or cynic. I don't see
many people asking for that one.
Critical and sarcastic or robot
efficient and blunt people complaining
about it being robotic. It can be more
robotic. Listener is thoughtful and
supportive. I think that's the closest
to the empathy people are looking for.
Although OpenAI has said they're working
to soften the overall profile of all of
these personalities in response to
customer feedback, the pitchforks or
nerd exploratory and enthusiastic. So
you can pick that personality. Now there
are other custom instructions and this
is what I've been saying when you
customize chat GPT take advantage of
that, right? Like I have it and I've
introduced it. I've said I'm Nate. This
is what I do and I've given it traits
like for me I want it to think strategy
first. I want it to be reflective. I
want it to focus on high signal. I want
it to push back on me. And so those are
things that I've actually put into the
customized instructions because they are
what I want. You can do what you want
with your custom instructions. And I
think people are sort of sleeping on
that as a way to handle chat GPT5
because that's what custom instructions
are for. That's that's exactly what we
should be doing. All right. So thinking
mode costs absolutely fixable. In fact,
I think that's one of the easier ones.
Guard rail friction is interesting. And
so there's there's certain cases where
you are going to have appropriate
questions for chat GPT 5 and it is a
little bit more conservative around dual
use content and there's particular risks
especially around biohazards that it's
super conservative about. Well, you
probably want to think about how you use
the model and how you ask for safe
completions in those cases. That's a
fairly limited like that's a very narrow
wedge, but it's something that comes up
if you were in biology, if you were in
research. You may be asking for things
that are entirely appropriate, but they
tend to be right next to things that
would be inappropriate to ask about. You
are going to essentially need to evolve
ways to talk to the model that
prioritize safe completions in ways that
are useful. Either that or you're going
to honestly have to switch models for
that one. Number nine, where it makes
basic errors, the simplest fix is to
require thinking mode. And the second
simplest fix is to require verification
and citations for factual claims. And
you can actually lean on that in the
custom instructions as well as a way of
reinforcing that. Now, I will say I've
emphasized customization and custom
instructions a lot before we get to the
10th thing here. It will not override
the system prompt in the chat if you are
using the chat. One of the ways you know
you can't overwrite it is you can demand
that the custom instructions be verbose
like super long- winded but OpenAI has
to preserve their GPU capacity and so
they are going to still impose token
constraints and you can actually see it
in the chain of thought. I've tried this
if you ask it to be verbose and write
long- winded stuff it is going to come
back and it's going to say I have to
respect OpenAI's token policies so I
have to watch my output length. It
literally shows you in the train of
thought where it's adhering to the
system prompt. And that's just good to
know because essentially OpenAI has put
some guard rails on that system prompt
so that you can actually not break their
GPU. I will do a separate video where
I'll break down the system prompt that
got leaked. I think it's super
interesting. It's too long for this
video. Uh but we'll get into it. It's a
super interesting system prompt. So
number 10, the silent fallback. So if
you are on one of the lower plans, not
one of the pro plans, if you hit
something like 80 messages in 3 hours,
it is going to silently downgrade the
model and the quality can drop mid
conversation. There will not be a
warning. The only solution here is to
monitor your usage and Chad GPT is
working on a way to monitor that because
they know that people want to see it. To
use the API if you care about that as a
developer or if you're not a developer
to upgrade tiers if you really really
care or to just go touch grass and take
a walk. I wish there was a way to force
it to sort of buy a prompt pack or or
buy an upgrade pack for three hours or
something. I think there'd be a lot of
interest in that. That is not something
that Jet GPT as a business has decided
to do.
All right. So, reviewing where we've
been going through these 10. Number one,
router misouting is a huge huge issue.
You can fix this with prompts like think
hard and also with custom instructions
like default to deep analysis. Number
two, chat and API being different
because chat GPT uses a routing system
in the chatbot and a API users can
select the model. Well, honestly, the
simplest fix there is to select the
model. Or if you are using the chatbot
on one of the higher tier plans, you can
actually drop down and hit the model and
like actually see like pro mode or
whatever you want to test. You can also
use the same number one fixes like think
hard if you don't want to go if you
don't have the option to go and hit the
drop down. Number three, model
retirement drift. So if a model was
retired and your old workflows broke,
what do you do? It's all about prompt
versioning and making targeted upgrades
and evaluating what happens. You should
already have prompt versioning and you
should already be evaluating. I've been
preaching that for a long time. If you
haven't been, this is where the bill
comes due. Please start now. Number
four, long context illusion. So, people
assume that because of the
advertisement-like quality of that
OpenAI live stream that they could stuff
in hundreds of thousands of tokens with
perfect recall, but that's not what
OpenAI actually claimed, and certainly
not what I'm seeing in practice. You
still need to use your good long context
practices like U-shaped prompting where
you emphasize at the beginning and the
end what you're looking for and
reiterating reiterating through the
context window what you want. Context
engineering still matters. I've been
saying for a long time there is no way
around good prompt engineering and good
context engineering. That is a durable
skill. Hey, it's a durable skill. Number
five, JSON breaking. It feels like a
narrow one, but it matters. We have had
issues with smaller models with JSON
breaking and not forming correct JSON.
Either upgrade to a better model or be
very clear that you want correctly
formed JSON in your custom instructions
and prompt for it very specifically with
like you want structured outputs in
complete JSON schema. Number six, tool
action claims that are not true, like
hallucinating tool calls. So this is
where I called out that with this model
in particular, getting artifacts
matters. It's a way of forcing the tool
call and forcing proof of tool call.
Number seven, thinking mode cause people
not wanting to use thinking mode when
they don't want to. That one is actually
one of the easiest. You just default to
non-thinking. And if you really want to
emphasize it, you can say don't think,
act now, or get the faster answer, which
is a little button that they added in
chat GPT. Number eight, guardrail
friction. This is another narrow one,
but it's for the bio researcher folks
out there, the folks using it for
science and hard science. You may be
asking queries that are close to
dangerous requests or requests that
OpenAI has deemed dangerous and it's
using safe completions. You need to
figure out how to narrowly tailor your
request in the prompt or you need to
switch models. Number nine, where it
makes basic errors is probably using the
non-reasoning model. So either upgrade
to a better model or adjust your
customization to require verification
and citations for factual claims and
really lean on that in the prompt as
well. Then number 10, the silent mini
fallback where like you use it 80
messages in 3 hours and it disappears. I
wish that I wish this was fixable but
like the open AI has to either up the
limits which historically they tend to
do or give you the ability to use a
different model which they've talked
about bringing 40 back or you're going
to have to monitor your usage and maybe
upgrade tiers. Now, there are people
when I say at the beginning of this
video, there are 10 things we can do to
fix these issues, who are going to throw
up their hands and say, "Why do I have
to fix it? I was promised a magic
thinking machine that would do the
routing for me and do the thinking for
me." I've seen that in my Tik Tok
comments over the weekend. I was
promised this and it didn't happen.
Guys, there's no such thing as a free
lunch. We spent an entire year asking
chat GPT to take away all of the other
models and give us one model that thinks
well. And people will say, well, we
didn't ask them to take away the models.
But most people did. They said they
didn't want the model drop down. If you
don't want the model drop down, you want
one model. Something has to give. And so
now we have one model in the drop down
or maybe like a couple flavors of the
same model in the drop down depending on
your plan. And we have to decide what to
do with it. And there is no way you can
make a transition that big and not have
some teething problems and not have some
issues with roll out and not have some
issues with how we learn to use it. The
idea that the intelligence from the sky,
the magic rocks that think are going to
magically be able to in a new model roll
out understand exactly what you want in
your vague English is not ever something
that you should have anticipated to be
blunt. It just isn't. Prompting is a
durable skill. Understanding how models
work is a durable skill. And
increasingly being able to adjust and
evolve your workflows with a new model
is a durable skill. That's not going to
go away. I'm going to keep exploring how
chat GPT5 works because this is a model
that is incredibly important in the
world right now because it's now the
only model that hundreds of millions of
people are using every week and it is a
complex model to use. My early
impressions are that this takes more
effort and more thinking and more
deliberate intention to use really
really well. even if the default feels
kind of smart to some people. So the
default may feel cold, but it can feel
kind of smart enough to some people and
I've seen that in my comments as well.
If you want to use it for extraordinary
work, which this model is capable of,
I've tested it. It does incredible work.
And I will do some more demos later this
week that sort of show that you need to
be ready to put in extra work versus
what you would have had to do with 03 or
with 03 Pro or with Cloud4 Opus. And you
might be like, is it worth it? Is the
extra work worth it? The answer is yes.
I have seen this model do analysis that
I haven't seen any other model complete
successfully. It can oneshot or fewot
coding examples for for software that
you can use around the office that I
haven't seen anything else do quite as
successfully. It is worth the effort,
but it is work. So, I hope this review
of 10 common issues across chat GPT5 has
been helpful. We continue our
exploration of this new model we're all
living with now. Uh, let me know what
you think in the comments. There'll be
no other issues I can address.