Everyone is getting the Opus 4.8 story
wrong. And I think it makes sense that
we're getting it wrong because we're
used to the 2025 story. The 2025 story
of AI was basically new model drops,
open AI drops, cloud drops, etc. And you
get a new high bar and then we talk
about what that enables, what that
unlocks, etc. We are in a different
stage of the race and it was never more
clear than when 4.8 dropped on Thursday,
May 28th. What happened was this opus
4.8
in some ways by some measures is the
strongest model out there right now. But
that doesn't mean anymore that it's the
best model or the most useful model for
you. And I want to unpack what's really
going on there, where I think it adds a
ton of value, where it doesn't, some of
the nuances people aren't talking about
enough that I think indicate where big
models are going, and really talk about
the state of the race between the two
major players left, open AI and
anthropic. So the first thing to
understand is that 4.8 8 is not the big
model drop that everyone is waiting for
from Anthropic. And we just have to get
the elephant out of the room. Everyone's
waiting for Mythos, right? Everyone is
excited about Mythos. Mythos is the most
teased model release in history. The
real reason is because they had a
funding announcement that they needed to
announce that day and they wanted to do
very classically a we release a new
model that's a leader in many things
plus we just raised a bunch of money.
That is why they released 4.8 when they
did. That was the reason for the
calendar timing. it is not because they
have the best new model out there to
drop. And I think that that distinction
came through in the test results. So
what you see is that 4.8 has made real
progress in some of these longer running
tasks that we know the model makers are
obsessed with because they burn tokens,
right? Like these longer running agentic
tasks. It does that well. It is better
at paying attention and staying on task
than 4.71, which is a weakness that I
noticed with 4.7.
That's great. But it is not the monster
intelligence super model like mythos
that everyone has been waiting for and
hoping for from them and that we all
suspect they cannot release because they
don't have the compute to deliver. And
so I think the right way to look at 4.8
is it's kind of like a placeholder
release. You need the release for your
funding announcement. You now have a
valuation close to a trillion dollars.
You've raised a lot of money and you
need to show that you are still in the
race and you release a strong checkpoint
of a model that shows that, right? That
shows that you're that you're making
progress and continuing to deliver and
everyone's still going to be waiting on
mythos. But even that, even that is not
the real story, folks. Because the real
story is that no matter how good it is,
it is not becoming my daily driver. And
it's not becoming a lot of folks daily
driver because of two key differences
that I think highlight some of the weird
dynamics we're getting into in 2026 with
very large models. And I want to be open
and honest about this because I think
that the way model releases and model
development is working in 2026, it's
really different from 2025. The first
big piece that I think is preventing
this from being a daily driver for a lot
of folks is it does not work predictably
when you scale up reasoning effort. And
so for I don't know over a year now we
have been told you scale up reasoning
effort and you get better results.
That's what everyone says. It might be
more expensive but you get the better
results. That appears to not be the case
predictably with 4.8. There are some
situations where scaling up the
reasoning level to what they call max is
going to be the best choice for you.
There are some situations where high
will be a better choice for you. And
that's super confusing because high is
less than max. And that's obviously less
than than the rest of the the reasoning
scale that we run into with OpenAI
models. Right now, if I scale up OpenAI
to the extra high reasoning mode, it
works better. It works predictably
better. That's nice for me because it's
it's really a product choice at that
point. I can understand how it works.
And it's not just me saying this based
on vibes, by the way. There are hard
test results that show this. Vending
Bench came out. Vending Bench is
famously the the benchmark that shows
how AI does at running an actual
business which is a vending machine.
Opus 4.7 did really well on this. Opus
4.8 did worse. It did worse people than
4.7. It was a regression. And that is
true whether you were on high or on max
for your thinking mode. And what's
really interesting is that 4.8 on high
beat 4.8 8 on mats for vending bench. In
other words, if you are running a
vending machine right now with your AI,
which I don't know a lot of people doing
in practice, but it's still a good
benchmark. I love the focus on practical
business, then you should be using 4.7.
It beats everything else out of the
water. And if you use 4.8, you should
use the dumber version of 4.8 because
max is not good. And I think that this
gets at one of the larger issues or
challenges with this current direction
from the anthropic team that I want us
to talk about more openly. 4.8 is a
model that thinks a lot about whether
something is aligned. And in principle,
you want more powerful models to be
aligned. I I get that. But if a model
overthinks, it may become less
effective. This ties in, by the way,
with one of the other big beats of the
last week or so, which is, of course,
one of anthropic's co-founders in the
Vatican in Rome when the Pope released
his encyclical about AI. Effectively,
the Pope picked a side with anthropic
and said, "These are the guys who are
thinking philosophically in a way that
is aligned to where I'm going."
Anthropic spends a lot of time thinking
about how to get AI right. And there is
a lot I admire about that. I have so
much respect for the work of Amanda
Ascal and the work she's done on the
constitution for Claude like it's
personable that feels like it has an
understanding of what it is to be human.
Now, I'm not saying it does, right? I'm
just saying that there is an ability to
grow a model because they're grown.
They're not they're not made that feels
like it understands the humanities and
that comes out in things like front-end
taste, which is really fuzzy, or the
ability to write sentences that don't
feel robotic, like all of that stuff.
But you can take that too far. And it
looks to me like 4.8 took it a little
bit too far because if you get into a
situation, and we have seen reasoning
traces that are coming out of 4.8 8. Now
on max mode, it looks very much like the
model overthinks itself and overthinks
itself specifically around the
constitutional questions. And what I
mean by that is what is right to say,
how do I align with my constitution,
etc. I I have seen reasoning traces
where where you just say something
fairly simple and you pull the reasoning
trace for 4.8 max and it's talking about
how it needs to write warm paragraphs.
It needs to align to its larger
constitutional questions. needs to be
aware of Amanda Ascll and sort of her
preferences, which is kind of funny
because I doubt that she would wish that
her particular preferences are recorded
in the model's thought patterns. It's
more that she happens to be a fairly
well-known personality now. She has
done, you know, larger conversations on
the internet about how she's shaped
Claude. And I think that there may be
some leakage from some of her public
statements on Claude into the model at
this point. We'll see. That's a that's a
suspicion. But regardless of how you
read it, 4.8 thinks to itself a lot to
the point where it is less effective.
And I think that that's something that
even if that's not true across the
board, because don't get me wrong, you
might think that I'm saying 4.8 isn't
good at things. No, it's very good at
some things, and I will get to that in a
second. It's that it's hard to predict
and reliably use it as a daily driver if
you have an overthinking problem. If if
it unpredictably overthinks about
things, can you trust it to be your
daily driver? And this brings me to the
second point that I want to make. Daily
drivers are increasingly a function of
harnesses. I talk a lot about harnesses.
I want to make it really simple and
clear. The harness is the shape of the
product around the model that allows you
to do useful things with the model. And
I'm going to be very specific and very
clear. In contrast, 5.5 in codeex with
4.8 in co-work and 4.8 in cloud code.
One of the things that enabled the
breakout in January is that claude code
was so ergonomic for developers. You
could type in plain text in the shell in
the terminal. You could get what you
wanted done. Claude would just
understand. Increasingly, it got into
sub aents. We got into the Ralph
looping. So, it would go on and do big
things. It's amazing how quickly things
change because that world that felt so
ergonomic in January and February hasn't
really changed. It's still what it was.
It's still beautiful in many ways. We
have of course continued development and
continued iteration from the other
player in the game from OpenAI. If you
are trying to tackle complex, difficult
work that is on the edge of model
capability, which is much farther along
than it was in January, I cannot
underline enough how much has moved
since January, how big the agent jump
has become. I know it does not show up
in the Excelss. I know it does not show
up in the PowerPoints. I know when I say
you can do much more with agents, people
roll their eyes and say, "Well, how is
that going to help my Tuesday?" But the
difference for anyone who is driving
these models at the edge is absolutely
stunning. And I I think it's really
important to be honest about that
because it shapes the ergonomics of our
workspaces which comes back to the
harness piece. Right? If we are looking
at harnesses and how they are useful in
late May, in June of 2026,
we have to recognize that the tasks we
are giving our models don't need a Ralph
loop anymore because the models just
know to keep going till they're done.
They don't need special help to stay on
task. They increasingly don't need
special help to review their work
because they have been trained by the
model makers who care about doing
longunning tasks well to get that piece
done. And so a lot of the things that we
associated with the harness, they've
evolved really significantly. And so
when we compare 4.8 versus 5.5 today, we
have to compare them for the task that
we're tackling them on. And honestly,
5.5 in codec is a much much stronger
harness right now, regardless of the
model inside, than 4.8. And I know that
in part because I've played with 4.8 and
5.5 in their respective harnesses. And I
can see places where 4.8 has more
insight, where 4.8 has more front-end
taste, where 4.8 is the better writer
out of the box. And I still find myself
going back to 5.5 in codeex. And the the
honest reason is the harness. And I'm
going to name specific aspects of this
harness. And I'm going to name them in
part because I want more competition in
the space. I'm not choosing OpenAI
because I am picking a favorite. I am
choosing OpenAI because behaviorally
that is what is working for me right
now. I think one of the key aspects to a
harness if you are going to do these big
long running tasks and for perspective
I'm running multiple two three four five
six hour tasks a day where I give the
model a big goal and it just goes and
does it there is no comparison right now
between where open AAI is at and where
anthropic is at on this and it's a big
big gap if you do big tasks and if
you're wondering Nate what are your big
tasks what how are you doing this I'll
give you an example
Yesterday, as part of my workup of 4.8
and 5.5, I gave both of them the task to
design end to end and build a website
for a MD Markdown domain that I happen
to own. And I was like, you should be
able to just do this, right? I shouldn't
have to remind you. I shouldn't have to
give you verifier steps. It shouldn't
just be a one-page website. I you should
be able to just do this. And what I
found was because of compute
availability 4.8 8 just errored out and
couldn't do more than than one task at a
time and took forever doing it. And what
I found with 5.5 is I could build two of
those sites at the same time. They built
relatively quickly. I did not love them,
but I had time to go back and I went to
chat GPT images mode and I actually had
said, "Look, I'm not happy with your
design initially, 5.5. I'm gonna ask
chatgbt images mode to design me a
better like JPEG frankly like a little
PNG image that shows the front page of
the site that doesn't suck. It is like
welldesigned and then I'm going to feed
that back to you and say look at this
image make it better. And I was able to
get through two websites and that
iteration loop that got to a nice fully
deployed on domain DNS name servers
assigned complete website with 5.5 twice
in the time it took for 4.8 8 to error
out twice. And I just I can't help with
that, right? And then there's smaller
things, right? If I tell 5.5, go look in
my files for X or Y, it just does it. It
knows my whole computer. It can sort it
out. If I tell 4.8 in the desktop app in
Mac to please go look at my files, it's
like, oh, I can only see downloads. I
can see desktop. That's it. and it
doesn't take the initiative to say, "Can
I get your permission to go look at
these other files since you clearly want
me to?" These are the little things that
make it hard to do big longunning tasks
if you are an AI builder in 2026. I know
it feels crazy, especially if you're a
CTO or a CIO where you're like, "Oh my
god, we we just did this. We just signed
the anthropic contract. I know people
who did this, right?" And they're just
like tearing their hair out. And I am
not saying that's a bad thing. One of
the things I think is really important
to recognize is that if you try to tie
an enormous amount of your budget to one
horse in this race, you are not setting
your company up for success. You should
be tying your budget to outcomes that
you want to drive. And you should be
allocating your budget against the
models that work best for those
outcomes. It's really very simple. And
so you should be in a position with your
harnesses, with your model where you can
swap them out when they don't work. And
it's just an API swap and that's it. and
you're done. But I really want to
encourage folks, don't pick a winner.
I'm not picking a winner. I'm not saying
OpenAI is always going to work best. I
don't know that. I don't assume that. In
fact, history suggests the opposite.
History suggests a continued horse race.
And soon I'll be talking about how
incredible Claude is because Mythos is
finally released or something like that.
And I'm open to that. I'm excited for
that. I like the story here. I like the
competition. I think it's good for all
of us. But right now, 4.8 8 is reading
like a checkpoint release that
overthinks itself that is unpredictable
and I think that illustration with 5.5
in the harness shows a lot of those
aspects right it like 5.5 was able to
get the files 5.5 was able to get the
entire job done twice in the time it
took 4.8 to think think and error out.
And these things matter. If I'm able to
run two or three or four or five or up
to 10 threads at once with 5.5 in
harness and I error out on one or two
with Claude if they're big tasks, I'm
sorry. It doesn't matter how smart your
model is. I can't pick it. But I said I
would tell you some of the things 4.8 is
good at. And I want to call one out here
that I think is really important.
workflows command is a really
interesting command in cloud code that
came out with the 4.8 release and I want
to name it because I think it shows us
an interesting direction for agents in
2026 and I think it's going to get
copied because sometimes these anthropic
innovations they just get copied. We had
a problem when we were running workflows
where you either deterministically tell
Claude this is the workflow I want you
to run. These are the sub agents I want
you to run. Uh or you let Claude sort of
decide how to get the job done. But you
don't get visibility. There's an in
between state that SLworkflows invokes
that's really interesting.
Slashworkflows as a command in claude
code lets you say please compose a
workflow and then claude 4.8 8 will
think through the problem, compose a
workflow with multiple agents, disclose
that workflow, and then give those sub
agents tasks in line with that dynamic
workflow. It gives you transparency. It
lets you see how agents are going to
tackle tasks that get the whole job
done. I think it's a pattern we will see
copied for a lot of individual
productivity agents in the summertime of
2026 because it just makes sense. Like
even with codeex, I can't do that right
now. That is a unique thing. I'm sure it
will get copied soon, but for now, it's
a unique thing with Opus 4.8. I think
it's a great innovation. Hats off to the
team. I love the idea. But this brings
me to one of the larger challenges in
covering models in 2026. And it's true
of 4.8. It will probably be true of
every model going forward. We are in a
world where agents are being used as a
single word for the productivity
enhancing agents that we have and also
for the larger team and org scale
workflows that we are building. And that
is really confusing, right? Because if
you're trying to understand how does
this make a difference for me, the
answer is going to kind of depend on
whether you're really in charge of
building those bigger pipelines at work.
It's like a an invoicing pipeline,
right? Something like that. Whether you
are in charge of enhancing your own
productivity on your team. And so
slashworkflows is not a command that
automatically works for these larger
pipelines, but it is a command that
works if you're enhancing your effort as
a claer, as a developer on your team.
Your individual productivity can
improve. One of the big questions of the
summer of 2026 is how do those two
pieces connect? If you are trying for
example to build an agentic production
pipeline for your engineering team, you
have to ask yourself how do you allocate
agents against a single source of truth
whether that's your ticketing system or
your repo in such a way that your
individual engineers are productive but
that individual productivity actually
layers up to something larger. And
that's where we see the larger unlocks
is if you can start to think about how
to write the entire rest of the pipeline
post initial engineering work so that it
is agent native first. That's when you
start to get the unlocks as opposed to
getting stuck in the sticky handoff. I
saw this happen when I was looking at
how Uber was complaining about their
token spend and there's been a bunch of
like highle folks who are leaking that
they're upset about their clawed token
spend. And the thing I called out is
that there's a big difference between
building agents for individual
productivity and building an agentic
pipeline that's native across the whole
system and ruthlessly hunting for human
handoffs that get in the way of
unlocking that downstream productivity.
Because right now we have the piling
problem. We have agents basically piling
up a lot of work downstream and if you
can't figure out how to manage that
bottleneck as it moves through the
system, you just have a giant pile for a
human to review at some point in the
system. There's there's no other way
around it because agents are so good at
generating stuff. And so when we talk
about workflows, when we talk about 4.8,
we need to understand that it is going
to accelerate that if you don't have an
agentic pipeline and and you need to
think about it very seriously if you're
in leadership, especially engineering
leadership. How do you create an agentic
pipeline that actually is more of a dark
factory approach? A dark factory, just
like a shorthand for you, is you submit
the PRs from your engineering team and
agents are the ones that handle the
merge conflicts. Agents are the ones
that handle the first, second, and third
PR reviews. Agents are the ones that
actually look through and monitor the
results in production. Agents are the
ones that review the the the other
agents work in this whole system.
Everything is identified because if it
is not, the work piles up unsustainably.
And that doesn't mean that humans aren't
involved. People think it does. It
doesn't. It means that humans are
increasingly over the loop. Not in the
loop, over the loop, monitoring,
designing the loop, and making it more
effective. SLworkflows
will just generate more downstream work
for you if you don't think that way. And
this brings me to the knowledge work
piece. If you are a knowledge worker, if
you are not a coder but like ending up
being code adjacent because that's where
we all are right now, then the best
thing you can do is to think about your
work kind of like that code stream and
think about these new model releases and
say, am I using these new model releases
to generate downstream work for my
colleagues in a way that's unsustainable
or am I using them in a thoughtful way
that enables me to accelerate toward
overall outcomes in the business?
because that question is going to be the
biggest question for the second half of
2026 for businesses. They want the
outcomes. So, how do you get ahead of
that and think that through? And when I
look at that and I back that into the
harness question that we've been talking
about throughout this video, one of the
things that's really compelling to me is
that the codeex harness is more
self-aware. And so, I can talk with
Codeex and I can say, "Help me think
through this outcome I want to drive
with my team. I want to set up
automations that enable me to do my work
more effectively without generating
unsustainable burdens. And Codex can
strategize with you about that. Codex
can think that through. Codex has the
ability to do computer use and the
ability to handle files in ways that
help you think that through. I'll give
you an example which requires one
computer use in a way that Anthropic has
struggled with. I know that the
anthropic computer use score for 4.8 is
very good. It's supposed to be better
than 5.5, but in practice, codeex's
harness with computer use actually works
quickly and works dependably. And that
makes all the difference in the world.
And I can ask it, use computer use, use
your in inbuilt codeex browser and
please set up this automation in such a
way that the output format is not
overwhelming to my colleagues who are
doing ticket triage. And it will do that
and it will give me the suggested
template and we can align on it and
we'll set up the automation. it will
execute it reliably on my computer even
when I'm away touching grass. Yes, I do
touch grass sometimes and claude is just
not there right now. And so my take for
you on Opus 4.8 is that you need to be
in a position with your AI work where
you are thinking more about the harness
than the model. And that's what this
whole video has been about. It's been
about the fact that 4.8 is very good and
I will just say it again, very good at
front-end design, very good at writing.
These are things that are classic
strengths of the opus lineage and of
claude more broadly. But you need to
think about your goals, what you want to
do, the outcomes you are driving and
back that into whether that makes it a
daily driver for you. And so if you are
a knowledge worker, I would increasingly
say ask yourself where are my outcomes
coming from? If I am someone who needs
writing help or who needs front-end
design help, I would increasingly say if
you're high volume, you're going to have
to use codecs and write fat skills that
cover those gaps or maybe work with Chad
GPT like I did for the website. If
you're not super high volume, then using
Claude can make a ton of sense because
it's just natively there and it's easier
to work with. If you are an engineer,
statistically speaking, and I've seen
these numbers in surveys, 70% of you
roughly are using cloud code and like 25
30% are using codeex and that number is
shifting around. And then there's like
this this like other section that's
there that's a bunch of open source
models. You should be looking at your
tooling and making sure that your
harness allows you to be productive in
line with the overall outcome of the
team, not just individually productive.
And that's one of the things I want to
call out with SLworkflows is it's an
incredible tool, but it's going to force
you to think about that sooner than
later if you're using cloud code, which
like statistically speaking, twothirds
of you are. And if you're using codeex,
you still have to think about outcomes,
but I love how self-aware the harness is
there. If you are a leader, if you're a
CTO, a CIO, and you're like, "Ah, Nate,
come on. You keep talking about this
Chad GPT thing. You were talking about
Claude before. I'm so tired." I've got
news for you. It is a two-horse race.
I'm not going to stop talking about the
fact that both of them have a lot of
strengths and that you should expect
another you know lead in the race from
for example Claude in in a couple
months. You should expect your system to
handle that. I will give you one more
juicy detail here. These 10 trillion
parameter models uh Mythos is in that
class 5.5 is in that class. Uh there are
others roughly speaking right some of
them they're not admitting it but you
can kind of tell if you use them a lot.
You should expect more open source 10
trillion parameter models by the end of
the year. And so you should be
architecting your system so you have the
option to hit very strong open source
models by the end of the year. It's just
at that point knowledge work will
largely be solved. So why not why would
you assume you have to spend it on a
particular model maker? Architect your
system for flexibility. So that's my
overall take on 4.8 very strong model.
It has a problem with consistency driven
by overthinking and it is it is not
fitting the harness as well as it
should. And codeex in 5.5 is fit they
fit handin glove from a harness
perspective and I think that's really
important to think about and I've tried
to give you specific examples so you can
actually see why harnesses matter so
much. I don't want harness to be a
foreign word. I want you to understand
it's all of the it's all of the the
scaffolding around the model so it can
do its job and as the models get
stronger you have to adjust your
scaffolding to work and that's part of
how codeex is strong right now is
they've adjusted their scaffolding so we
will stay tuned I'm excited for mythos
if you want to get the full breakout of
the tests that I ran for this head to
the substack if you want a great great
guide for which you should pick up for
which thing because I've given you
general principles here but like you
should dive in and actually figure out
which works for you, how to get started.
I have specific guides for you on
Substack for both 4.8 and when to use
max and when to use high and also for
codecs in 5.5 so you can compare them
and figure out what's best for you. And
yes, you can obviously feed this to your
model to figure that out as well. And
that's one of the things that I put in
the substack is like a guide that you
can like feed the feed feed the thing
in, have the conversation, figure out
what's best for you in a conversation
because so much of what we learn is
conversation. Okay, I will catch you
next time. 4.8 is one of the most
interesting model releases I've run
across and I I think it illustrates
where we are in the race.