ChatGPT 4.5: Expensive Strategic Lego Block
Key Points
- ChatGPT 4.5 launched today with substantially higher pricing – about $150 / M tokens for output and $75 / M tokens for input – roughly 10‑25× more than Anthropic’s Claude 3.7 Sonet, making it cost‑prohibitive for most users.
- Because of the massive compute needed, OpenAI limits 4.5 to Pro‑plan customers for now, and even announced a need for “tens of thousands of GPUs,” a move that coincided with a noticeable dip in Nvidia’s share price.
- The pricing and compute surge reflect a strategic difference: Claude positions itself as a specialist (especially for code) and can afford a narrower focus, while OpenAI must keep ChatGPT as a universal market leader that covers every use case.
- OpenAI’s roadmap for 4.5 emphasizes “non‑benchmark” capabilities—emotional intelligence, nuanced style, and surprising creativity—that improve real‑world user experience and are intended to become core building blocks for future models.
- The speaker argues that 4.5 should be seen as the last “Lego block” before a next‑generation, hybrid model (e.g., a GPT‑5 slated for Q2) that combines lower‑cost inference, reasoning, and the new emotional‑intelligence features into a more compelling, sticky product.
Full Transcript
# ChatGPT 4.5: Expensive Strategic Lego Block **Source:** [https://www.youtube.com/watch?v=KeBy7imDM1A](https://www.youtube.com/watch?v=KeBy7imDM1A) **Duration:** 00:05:26 ## Summary - ChatGPT 4.5 launched today with substantially higher pricing – about $150 / M tokens for output and $75 / M tokens for input – roughly 10‑25× more than Anthropic’s Claude 3.7 Sonet, making it cost‑prohibitive for most users. - Because of the massive compute needed, OpenAI limits 4.5 to Pro‑plan customers for now, and even announced a need for “tens of thousands of GPUs,” a move that coincided with a noticeable dip in Nvidia’s share price. - The pricing and compute surge reflect a strategic difference: Claude positions itself as a specialist (especially for code) and can afford a narrower focus, while OpenAI must keep ChatGPT as a universal market leader that covers every use case. - OpenAI’s roadmap for 4.5 emphasizes “non‑benchmark” capabilities—emotional intelligence, nuanced style, and surprising creativity—that improve real‑world user experience and are intended to become core building blocks for future models. - The speaker argues that 4.5 should be seen as the last “Lego block” before a next‑generation, hybrid model (e.g., a GPT‑5 slated for Q2) that combines lower‑cost inference, reasoning, and the new emotional‑intelligence features into a more compelling, sticky product. ## Sections - [00:00:00](https://www.youtube.com/watch?v=KeBy7imDM1A&t=0s) **ChatGPT 4.5 Pricing Strategy Explained** - The speaker breaks down the steep token costs of the newly released ChatGPT 4.5, compares them to Claude 3.7 Sonnet, explains why it’s limited to Pro users, and discusses the massive compute investment and market positioning behind the model. ## Full Transcript
Chad GPT 4.5 dropped today like an hour
or so ago and we're going to talk about
the strategy because a lot of people are
confused and frankly they're confused
for good reason to start with 4.5 is
expensive and I can put dollars on that
because they price it per million tokens
Claude 3.7 Sonet which is another model
that dropped very recently it dropped
like three days ago it comes in at an
output cost of $115 per million tokens
and an input cost like sending something
in at three bucks per million
tokens by comparison chat GPT 4.5 which
dropped today output cost is 10 times
more
$150 per million tokens the input cost
is $75 per million token Which is vastly
higher than three bucks it's huge the
input cost is 25 times
more the higher computational costs are
real it's so real that Sam Alman could
not release this to anyone except Pro
Plan users right now plus is going to
have to wait apparently they're adding
tens of thousands of gpus which makes it
really funny that Nvidia fell like eight
or 10% or whatever it was today because
like he's literally talking about how
much compute he has to add to serve this
model and people are like why would you
put all this work in to a more expensive
model
when it doesn't
reason because 01 Pro reasons 03 reasons
Claude 3.7 Sonet is this hybridized
model it reasons it doesn't reason
depending on what you need it's focused
on code which is a high value use
case I'll tell you why the play here is
a legol block play chat GPT is a market
leader it is not a challenger Claude is
a challenger Claude needs to specialize
Claude is specializing in code chat GPT
is a market leader and needs to cover
all the bases to lead the market that
means they cannot Just Produce deep
research they cannot Just Produce 01 Pro
for inference and win they have to
produce a model that does everything to
earn the user base they have which is
the only user base in the hundreds of
millions they're the only ones and so
they have to do everything well and what
this is designed to do well is new nuan
stuff that isn't captured on benchmarks
but which chat GPT thinks is a long-term
building block to their success they are
highlighting emotional intelligence they
are highlighting nuanced writing style
the ability to surprise you these are
things that don't show up in an aim eval
but they do show up in real world
interactions for users and the long-term
bet is that they can bring the compute
cost down they can hybridize this with
the other models that they already have
in the stable
and they can produce a gp5 by Q2 that
has emotional intelligence built in
thanks 4.5 and has the other pieces as
well has the reasoning piece has all
this other stuff and so if you're
judging 4.5 by what is released today
you are probably not judging it
correctly you need to look at chat GPT
4.5 as the last Lego block in place to
build something that is much more
compelling and sticky as a customer
experience for chat GPT long term and so
chat GPT
567 whatever that's going to be is
dependent on getting these complex
Primitives right and arguably from the
compute cost emotional intelligence
nuance and the ability to surprise you
is extremely compute intensive and that
doesn't really shock me these seem like
really hard things for a machine to do
if a machine can do this very well
that's a really big deal
that is genuinely novel that is an
achievement that is really significant
even if it's hard to measure and so that
is what Sam is doing with GPT
4.5 it exemplifies yet again why it's
important to have real world evaluations
and real world conversations about
performance and capabilities because
these benchmarks are just not good
enough these benchmarks don't tell us
these things and so we're going to have
to all get used to this another good
example is Cloud 3.7 people are talking
about the fact that it is built to be
more opinionated with code that is a
designed decision it is a designed
decision that does not show up on evals
but it's really important and you can
disagree with it you can say you want a
more malleable model and you really want
to use 3.5 sonnet or you can agree with
it and say I like the structure this
provides I like that it insists on a
particular sort of way of building code
and I think that that helps me to build
quicker because the scaffolding is in
place you can have opinions but you need
to know what the model does to have
those opinions and we do not have other
than like digging in and talking about
it in places like this good ways to do
that we need better evals anyway that's
GPT 4.5 that's the strategy that's
what's coming you try it out if you're
in the Pro Plan and let me know it's
only in the Pro Plan right now it's
coming to the plus plan next week cheers