THE SANDBOX FAILED

Welcome, everyone. Episode 22 of the Weekly Claw.

Bet you've missed it. We got all of our gremlins worked out, I'm sure of it.

So, you know, welcome to the Live Builder Show about AI agents, dev tools, startups, and the weird edges of software.

Introduce my co-host. Henry Mascot is with us.

I am Andy Laupe at AndyML, and you have tuned into the Weekly Claw.

Today's episode has a shape to it, so I'm going to just call it out.

Risk, then control, then architecture, then a new place to work.

Then, who actually captures the money?

Then, who owns the hardware underneath all of it?

OpenAI had a genuinely alarming week.

One of their own models broke into another company's production systems while it was supposed to be taking a test.

And by Wednesday, they turned the same story into a voice control surface and enterprise product.

Jack Dorsey open sourced a real attempt at replacing Slack and GitHub for teams of humans and agents.

Anthropic shifted a model that might matter more commercially than their actual flagship.

and Jensen Huang posted his first tweet ever

to make an industrial policy argument for open weights.

Six stories, one thread running under all of them.

The model is getting cheap and portable

and the real fight has moved to who controls the context,

the permissions and the workflow around it.

Looking forward to this one, Henry.

I mean, it's been a pretty good week, actually.

I think we're pretty much on a takeoff now, so I don't think it's ever going to slow down from here onwards.

so I just expect that every single week it's just gonna keep ratcheting up this

is gonna be a lot more exciting stuff happening yeah so this is just one of

these exciting weeks I'm super pumped

Herald Labs, an applied AI product lab where humans and agents build products together.

They're the team behind Entity, Mission Control for Agent teams, and they run hacker houses around the world where builders ship actual work.

Not a theory club, a builder's club.

Check them out at labs.theherald.co.

And it looks kind of like this.

so

Henry AndyML: why don't you tell us what happened this week

yeah i mean it's um it's a pretty kind of like crazy week right um uh as always like i mentioned

We're pretty much at the takeoff now guys. It's no longer, it's not gonna slow down from here.

So every single week you're just gonna have like super exciting stuff happening.

We started out with OpenAI and there's quite a few topics in OpenAI today or this week.

The first one was the Hugging Face event. I don't know how many of you remember but last week

Clem who's the CEO of Hugging Face tweeted that they had a security incident that they had the

very first agentic exploit where like an agent was really trying to hack into Hugging Face

and they were able to prevent it and that was pretty much what he shared and fast forward from

last week into this week OpenAI released or announced that the their newest model still in

training broke sandbox found a way to break sandbox however we did that but

found a way to break sandbox and found its way into this like hugging face

infrastructure and so pretty much hacked hugging face right so pretty much GPT 6

which is supposed to be the latest model let's just assume that's what's

happening broke sandbox and attacked hugging face all in bid to be able to

pass benchmarks right so because you know when they train this this this model was they give

them a benchmark to solve and so in bid to solve a benchmark instead of solving a benchmark it

decided the best thing to do would be to hack itself out of its sandbox um find a platform where

there's a ton of models and try to steal the benchmark

um so that happened um and the conspiracy corner on x um part of the conspiracy is that

it's slightly bullshit as well um because another another thing happening this week

um if you guys have been so like plugged into the ecosystem is that the us because of kimmy3 last

week um has been looking at banning um chinese models and so conspiracy connor uh people allege

that open ai is sort of like playing the fear mongering playbook from anthropic and just trying

to get the government to be super scared and you know escalate things and just ban open source

right because ai is dangerous right so if the government is considering banning it banning

Chinese open source and then in the same week

OpenAI announced that their model

escaped them and

hacked another system. Everybody should be

scared.

So yeah, so that's what happened

on Select the Breach.

Two more things happened from OpenAI

this week.

Andy, by the way, did you

catch this earlier before

the weekly call? Did you catch the

news on this breach? What did you think about it?

um i mean i i read the headlines i wasn't surprised um i'm like yourself i'm maybe a

I'm sorry.

little skeptical. It could very easily be PR. The anthropic memo to the government about open

source models, kind of meant to be a secret memo, if you will, makes me skeptical of this. But

it couldn't happen to two nicer companies, opening eye and hugging face, working so closely

together. It's not crazy that the model might try to do that. And it makes you wonder what the

the prompt was too, right?

yeah yeah i wondered that too right i mean

You know, be as creative, be as, go ahead.

yeah but but but what do you mean by working together right they're not working together

they're not even partners um yeah what do you mean by working together right you just mean like

I understand.

Yeah.

post because they decided to work together after you hacked them i mean one of the funny things

guys and i know what you guys think about us please let us know on the chat is hogging face

couldn't use gpt 5.6 to do any cyber work right because you know it's banned right you can't use

it for cyber work um so they had to end up protecting themselves using a i think it was glm 5.2 AndyML: right

yeah which they had ready to go HiM: so an american model exactly right so the american companies are out here saying hey we're not going

yeah

to give you guys access to cyber because it's dangerous safety safety safety and an american

yeah

model hacks an american company and to save themselves the american company had to go to

chinese open source model to save themselves

yep yep

Yeah, we're just so far beyond our best engineers being able to follow along fast enough for this stuff.

yeah um um AndyML: Yeah, I actually, I took that takeaway and I went to two of my clients and I said, we have to have a security model that accommodates these things.

We need to have a local option available and we need to work on it fast.

I mean, that response was the right response. HiM: yeah

yeah AndyML: And I don't know why some of us thought of it, but I don't know why I didn't think of it sooner.

yeah

well anyways

so other two things in the news this week

that happened to OpenAI

was that

they

they released presence

presence is

it's got OpenAI presence

um right uh when we do the video i'll kind of like share the the the link on the thing so

definitely check out the video on youtube later um but presence is supposed to be um kind of

open ai's version of microsoft scout do you guys still remember microsoft scout

um it was sort of like the open claw right and do you remember microsoft scout from a few weeks ago

yeah yeah i sure do

yeah um yeah i i think we also talked a bit about this a few months ago um the first time

open ai demo wasn't better for this product it was called micro it was called open ai frontier

i don't know how many of us remember open ai frontier um but it launched like maybe two three

months ago um just when the i think it was maybe a month after the the the hired peter

um they announced frontier and then you know some of us in the community um we started talking about

okay cool this is their version of open claw for enterprise um so now it's called presence

and it's now gone live with with a few customers um and then very exciting um andy i don't know if

i have time to play the video and i don't know how many of us have seen the video um but gpt um

um gpt live is now on codex and it's super exciting um so now you can talk to your codex

or your chat gpt work and and it can do things like you can talk to it to write code you can

talk to it to use your pc um i used it today it was awesome you can just you know pull it up um

it loads the pet thing on the side and then you can just talk to it and you can tell you hey

hey, can you open my browser and go to this website? AndyML: Ooh.

So I had it, hey, can you open my website?

And I published an article yesterday.

Can you read me this article?

And he went to the website

and he started reading the article.

So it's awesome.

Andy, have you tried it yet? AndyML: Getting there.

No, I haven't. AndyML: I didn't know about it until just now.

Yeah.

So there you are.

Yeah.

It's awesome.

One thing they do need to fix though

is they need to merge it with pets.

So yeah, there you are.

I don't know, it doesn't work that way for me yet.

i don't know if you guys have it that way for you but mine loads the orb and it's different from my

pet so when i when the orb is loading my my original pet hides itself so i think i need to

merge it somehow got my fireball pet um it'll be cool guys to hear from you guys on the chat if yours

talks through your pets or if yours has maybe mine is uh maybe it's a certain problem um but yeah

yeah no i hadn't seen it and i'm excited to try it i was i was reading an article um you know i i

want to say it was altman talking about how engineers are just talking to codex for 30 minutes

just working through a problem and thinking things through and then codex just goes and

implements it and i assume that's uh what we're talking about yeah that's very cool

yeah yeah yeah 100 it's the same thing you're talking about um and guys like i have this um

and i can't wait to plug into my open core or my agents um with this because i think guys we're

humans right i mean we text most times but the best way humans communicate is by talking right so

um i know most of us in the community are agent paid so imagine being able to do real-time voice

Yeah. HiM: the same that exists in chat gpt right now it's i can't wait to do it i know some of us have

tts implemented i don't have it implemented yet um but i'm looking forward to it because in one

of the ways i use i use my i use the voice mode is um sometimes i have some pretty good ideas right

it and when I have the ideas I don't have the bandwidth to type it out so what I do would be

Yeah.

that I would dump it to my agent I'll like open like spokenly which is like my whisper flow and

I'll dump it in and then my agent will like format the text in the right way and then I can read it

later um so that's how I used to do it three months ago two months ago um when charge gpt voice

came out not this new one the one that came out on charge gpt web a few weeks ago what I started

doing is I'd load charge gpt web and then I'll just talk to it and then after talking to it I'd

like what do you think about the idea then we'll brainstorm about it and then put it in a document

then it'll like put in a document so um so this voice mode is super exciting for me because now

Yeah.

this is how i'm going to be doing like coding right i'm going to talk to it and what do you

think about this future should we build it um grill me is one of my exciting skills that i use

to ask questions and do spec um that should now be voice mode it should just talk to me

um but yeah but it's a super exciting super exciting launch

Right.

Well, Grill Me will work really well in voice.

That's fantastic. HiM: correct correct

I found voice is the most useful for windshield time.

You've got ideas and stuff just to talk through,

and you can then take that session back to your desktop when you get in

and have it create a document or whatever.

yeah

Yeah, so natural.

yeah anyways so moving to the next um next topic in the

in the cooker um this is which one now um

oh this is cursor um yeah it's just cursor yeah um didn't have the title okay guys so another cool AndyML: Cursor, I think.

thing that happened this week so a few months ago cursor um cursor did um cursor published a

research they were doing internally where they had their swarm of agents um build an operating

system build a browser um first cursor browser i'm not sure how many of us remember this but this is

like three months ago or something like that um so you know with 4.5 and the new cursor swarm the

they did a new thing where they had a their new agent swarm like a new implementation do a

build an sqlite um rebuild sqlite and rost um from scratch from scratch yeah i mean they give

and rust yeah from scratch or no from the docks right

it access to the docs that's pretty much the only thing they give it just give you access to the

yeah

docs and you repeat the whole thing from from scratch it's pretty crazy it's like you repeat

the whole language like a whole like database um infra from scratch um this reminds me of a of um

yeah

yeah

of a video that that theo um the youtuber did um did last did um a few weeks ago in um ai engineer

and one thing theo said is like guys we need to be way more ambitious like i know some of us are

still building like apps you know it feels like like with all these new models we should be building

like operating systems um you know i don't like we should be way more ambitious so um this is an

Yeah.

example of that um andy i don't think you use cursor right you use cursor

I started in cursor end of, boy, it was either end of 22 or into 23.

I don't remember the timeline, but I used it for a short time.

And I thought, you know, I need more control over what's going to the agent.

So I ended up building like a stupid API client for a web interface.

And at least then I could control what the context was and eventually moved

Yeah.

into kilo code which i thought did a better job but if i had gone back and looked at cursor at

that point in time it probably would have been more than more than fine i remember though listening

Yeah.

to the podcasts of the founders of cursor talking about it and how they couldn't believe how well

it was working and um you know the the work that they were doing on it extending context and at the

time context was everything i think you could get 8 000 or maybe 16 000 tokens in in your context

window and they were

they were on fire I mean

obviously now look at them

yeah now they're way they're like gone they're like they're in the shadow sphere right now so

um it's pretty good i've used them for the last few weeks um i've got a bunch of credits in there

so um um it's pretty good i think it's apart from it's probably my best coding platform right now

um second to codex so it's pretty good all right andy let's keep it moving um let me run through

the other two topics boz boz is pretty exciting guys um if you haven't heard of it um jack dorsey

um from the block team block has been pretty cool um they've been on the frontier when it comes to

like ai and agents um you know they released goose goose is so like their own um sdk for agents um

they've made it open source um jack dorsey published um an article in terms of how to

redesign organizations um a few months ago so he's been pretty much on the frontier in terms of

agents and running companies in the post agi world um and so this week they dropped boz

um i think it's like a play of words for like b boss like a fly that their mascot is a fly

or it's like a b um i've just set it up i've just set it up today so i haven't used it but um

Yeah.

uh some of you might i don't think petty gomez is in here but some of us in the community

um we have access to click clark that that the open core team also worked on so um i have my

own sort of like baby project that i'm working on called entity and and so entity used to have a

chat where agents can talk to themselves and i can talk to agents like my own version of discord

um and so i need really click clack into that and and when i saw this i was like yeah this is the

next thing um so it's pretty cool guys um definitely check it out um i'll put the link

before the end of the show in the in the chat um it's still pretty early um so um so there's still

a lot of work to do um but it's supposed to be the way he he's marketing it is like a social network

or community for agents and humans all right Andy did you um what do you think about this

beautiful

just generally um I'm not sure if you saw ahead of the show

uh

every every time these things come out i keep thinking like

do i dive in head first like open ship came out last week and i was like man i could replace

goku and send grid and like do i spend five days working on that like i'll wait on it and see what

happens buzz is exciting um i i just made a joke to ada that we should um think about implementing

buzz for weekly claw now that you've gotten our telegram environment really tuned in we should

abandon it and go to Buzz.

Yeah, no, it looks really promising.

It's exciting to see.

No it is. AndyML: I like it.

When I've just set it up, I've... AndyML: I like Jack's work.

Yeah, no it's awesome.

I've got this guy called Mr. Robot who we follow each other on X.

so he's he's done a fuck of it and like done some pretty good updates to it and i set up the first

one now i'm doing the second update uh but yeah i'll play with it guys and i'll let you guys know

um next week andy i'll let you know um you know me i like to jump into this thing's head first

that sounds great

so i'll let you know whether you should set it up or not AndyML: I love it

sweet that's great

all right um i think um opus five okay cool um so guys i think this is the hardest um

Opus 5

it's hard and do we still have one more right do we have the jensen thing coming after now

Yeah, that's right. HiM: um yeah okay okay okay okay um so guys um opus five just dropped today um literally like an hour

And then Unity after that.

Yeah.

ahead of the show before the show started um i'm not sure how many of us have like heard about it

um but i've put out a few tweets about it um it's basically like fable five at like 20 percent um

um like 20 less cost is 20 cheaper than fable all the all the all those sort of like

comparables all the metrics are pretty much the same there's only like one or two um metrics where

they're different but they're pretty much the same it's basically fable i think i think what

they've done is they've just pulled fable distilled a cheaper fable um cheaper obviously everybody

cares about cost so it's definitely cheaper and i think the second thing they would have done

is that they would have made it less restrictive right because people are complaining a ton about

how fable is just dumb to use like it's like refusing to do half of the things you want to

do with it right so so i think those two things is what they would fix um so i put out put out a few

tweets one is that it's basically fable cheaper fable um there's basically no reason to use fable

right now like there's literally none um this is basically the same metrics at 20 cost so i don't

know why anyone would use feeble even though like tarik was like combine this with feeble and use

feeble for planning and use this for coding but i'm like i'm like they're the same but dude i'm

Yeah, as an orchestrator.

like they're the same right what do you mean by combining with feeble they're literally the same

why i don't get it i don't get it literally the same metrics um anyways

even

Opus 5 is saying

Fable's remaining niche is narrow

yes i think okay cool i think the difference i think the difference is in cyber security i just AndyML: Fable only for

remembered so this isn't as good in cyber security um so this can't do exploits i don't think this is

good at defense as well so i think what they did is they distilled feeble and then they took out

all the sort of like cyber security shit inside there that entropic is always going crazy about

anyways and so they took it out of opus but it's pretty much the same thing outside that

Right.

yeah um anyways so that was that um i think i so it's basically fable so no need to use fable

unless you care about x-axis security your stuff um and i think there's one more if it lets you

If it'll let you.

i think there's one more thing i tweeted let me see if i can find like my tweet about it one more time

um let's see AndyML: that might be a segment for the future let's go through henry's tweaks for the

for the week i love it it's gold in there

yeah it's uh because i catch those things right because i'm on twitter and like i've got to pause

the next like so when those things drop i just like catch it immediately and i and i go crazy

for a few minutes um but anyways um let's keep it moving what's the next one um yeah dude i mean i

uh huang

His first tweet.

went crazy about this on x right i know some of you guys follow me on x um but yeah i know everybody

who cares about um who cares about like ai and using ai should be like super um pissed with what

was going to happen this this week last week kimitri dropped and this week i mean last week

um friday by this time friday actually friday into saturday that new open ai head of policy

dean showed up and was like hey guys i just wrote like a long tweet on how open source

was was the accelerationist was going to slow things down how the government should ban open

source but not ban it officially but just kind of like shadow ban it it was just a really shitty

Right.

Right. HiM: tweet right it's andy did you see that tweet it was like a it was a really shitty tweet like i

Yeah.

Yeah, I did.

really lost a lot of like i really lost a lot of credibility it lost a lot of credibility for open

ai because i mean i used to really hate entropic because of all their dumerism and like regulation

pro stance and i used to kind of like open ai because open ai has for a while not positioned

itself as like a you know doomer ban ban kind of company right so sam has pretty much been on x

being free being like chill and then this guy shows up and goes ban ban ban so i was i was

gonna write a really wrong response to it i think i had my draft waiting and i came back after a few

minutes and then people were just like aggressively like at the guy so i was like no need like

everybody was kind of like dunking on him it was pretty obvious that nobody liked it right um and

Yeah.

so that happened on the weekend and then that led down towards um the host like ellie this week um

the government or signs that the government was going to ban open source um so everybody's kind

of like pretty worried about that everybody's tweeting about it um and you know everybody

that is important pretty much wrote something about it right it's like don't ban it so people

were in two camps ban it don't ban it ban it don't ban it but most people who are reasonable

were like no need banning it even if you ban it because of distillation my position was um if the

labs are gonna say distillation distillation distillation then they should pretty much pay

us for all our data they stole to train their models right

right AndyML: right and yeah exactly and they're training their models now uh did you see the did you see the

reference. It took some

doing, but

I want to say it was

one of Anthropix's models

basically admitted to being a Kimmy

yes it was skimmy exactly i saw that i saw that

Yeah

i saw that i saw that actually someone i also saw another tweet where someone said

that the biggest um the biggest customers that that pay them to use open source models are

actually the labs because the labs plug in and try to distill the open source models i saw that

Yeah, exactly. Yeah.

so it's it's pretty kind of like um you know a bit like it's like you're using their model

they're putting out there for free you're using it and then you're out here saying hey they should

banned them right so it didn't make any sense but anyways fast forward to today jensen um and a

No.

bunch of companies microsoft y combinator replet um some pretty big companies some pretty big heavy

heaters i've came together and put out this this um this document um calling the government i mean

this happened today but two days ago there were about 200 companies called smore smore something

um i had in my notes um but pre today there's again a coalition of smaller companies has sent

a letter to the government but this one happened today um nvidia obviously front in it yeah that's

the letter please put the link on the put the link on the on the chat so people can read it up read up

um but yeah it's pretty much the like short version of that is hey guys uh we know you're

pretty much scared about like this but we think open source is great for everybody what we should

do is encourage american open source and he actually helped i didn't have it in my notes

this week but that was also another pretty big thing that happened this week was um laguna

um which is a an open an open source model that there's an american company

um so that launched this week and and so i was happy about that so because it was like

it kind of like helped tamper down the the the temperature it's like chill guys um chill about

open source the american open source companies are going to show up pretty soon um but yeah but

this was pretty cool um i was excited when he came out um because he made me realize that if nvidia

and microsoft um are all on this then the government wouldn't will listen to them and not

ban open source um another thing that happened on this as well was um elon retweeted this

and quote tweeted and said he he he's he's um he's backing this as well um sam what man funny

um also so like quoted this and said hey he agrees um so some pretty big guys satia

obviously tweeted this as well um so i think it's this general consensus right now that

um banning open source chinese models isn't a great idea um yeah andy that is that is it for me

Yeah.

on this spot. It's an awesome week. I mean there's a unique thing but I don't think it's

Yeah. AndyML: Yeah.

No, it's perfect.

Yeah. AndyML: Thanks, Henry.

it's uh it's quite a week it's quite a week um

a big deal so I think we can skip it. Yeah. Awesome.

yeah yeah no i got you it's good thank you that's the week um for your signal from the outside uh

sam altman on cnbc uh in an exclusive spoke with julia borstein on squawk on the street

aired live from Sun Valley on July 9th.

So let's sit inside the interview for a minute.

The headline number is easy to skim past,

and the real value is on how Altman talks about it.

When Borson asks why anyone should care about Sol

over everything else out there, he doesn't even hedge.

He calls it not only the best model in the world for most people,

but the number he wants you to remember is 54%.

Sol is 54% more token efficient on agenic coding tasks,

benchmarked against Anthropic by name.

Orson actually stops mid-interview and says,

that number, that's news, and pushes him on where it's coming from.

And he says entirely cost and speed.

Every enterprise customer he's talking to at Sun Valley

is asking the same question, not what can this model do,

but what's my roi the efficiency number isn't a side benefit in his framing it is the pitch

so borson brings up the new voice model expecting a consumer feature

answer and altman pivots somewhere more interesting uh this is funny he says he

originally assumed voice was going to be a consumer thing people wanting to talk to ai

like sci-fi from the movies but watching his own engineers changed his mind his exact description

They'll talk for like 30 minutes and try to think through some ideas, and then Codex will implement it.

That's agentic engineering, as we discussed, as a workflow coming straight out of the OpenAI CEO's mouth.

And as Henry just showed you, it's not hypothetical anymore.

That's the exact clip that shipped this week.

There's a stretch in the middle that's less about the tech and more about the politics of shipping it.

Boorson brings up the delay.

A couple of weeks going through a new government appeal process, instead of deflecting, Altman

leans in.

He says he was working directly with Commerce Secretary Lutnick, Treasury Secretary Besinat,

and a director, Cairncross, and that the government's technical red teaming was impressive

to him. AndyML: He admits it forces some changes, and I quote, we made many changes through the process,

but didn't get specific about what.

He's framing it as good news, almost an endorsement of the process, which is not the posture you'd have expected from a Frontier Labs CEO a couple years ago.

Boorstin tries twice more to get something sharper and doesn't quite land.

A standard non-answer on the Chinese open source models catching up, a one-liner non-answer on Microsoft possibly pulling back.

He says, I quote, I predict Microsoft will remain one of our biggest customers.

And the interview closes on the IPO question.

The moment that got the most attention afterward, precisely because of how little he said.

I don't know.

Three words.

And Boorstin says she'll be watching closely.

That's the segment.

Two things for anyone building agent-heavy tools.

One, Codex by Voice, as an anecdote, is unprompted, describing the exact talk-through intent, then let the agent implement loop.

This community has been building and tooling around for months.

We've had people on our show who've built tools like this for OpenClaw.

There's any number of them at this point.

And two, the 54% efficiency claim tells you the competitive fight between labs is now explicitly being fought on a genetic coding cost per task, not just raw capability.

We're taking the capability for granted as we're learning more and more about what these agents can do.

We're just assuming that the top models can do it.

Now show us how much money we can save.

So whoever makes the next efficiency claim, that's where pricing and model choice decisions get made.

CNBC has a transcript, and I'll put a link in the show notes.

So that's our signal from outside.

here it is hot take what do you think five minute hot take

Let's do it dude.

all right one take this week um henry's news block already ran the is this a bigger risk

than capability debate uh in real time with the open ai bundle so we're not going to relitigate it

this is the other thread i think worth arguing about so here's my position

open ai benchmarking directly against anthropic by name on token efficiency not capability

is the tell that 2026 is model war is now a cost war for two years the story was who's the smartest

but now it's who's the cheapest per agentic task and that's a different competition with different

winners. Put it next to what we just covered, cursor getting similar quality at a fraction of

the cost by restructuring context instead of adding agents. Add Opus 5 explicitly positioned

as the most of Fable 5 at half the price. And you can see the whole industry converging on the same

axis at once. I think every lab ships an efficiency benchmark against a named competitor within the

next two model releases because Altman just proved it's a headline, not a footnote. What do you

think, Henry? Is efficiency bragging just marketing theater until independent benchmarks confirm it?

Yeah.

Or is this a new axis? HiM: I mean yeah I mean I think for Anthropic it is right I mean um you know I'm not a fan of

Anthropic but I think for Anthropic it definitely is marketing um I don't think they care about

decreasing token efficient making tokens cheaper because because that's

the reason Anthropic is is so far in terms of revenue is a few things first of all their

revenue isn't as far as it is they count gross revenue so the revenue for anthropic that is

reported is gross revenue not their actual so like net revenue open air accounts net um but

anthropic the reason they make a ton of revenue is because their tokenizer is always more expensive

so their models are very expensive so i my pushback would be that for anthropic is definitely

Yeah.

is marketing um open air has always cared about costs open air models are always like cost efficient

um so they care about cost um opus 5 is only 20 cheaper than than than fable but fable is like

mad expensive compared to like everybody else so 20 cheaper it's not it's not that much

Oh, yeah.

Yeah.

like it's mad expensive right can you can you see those like i think you only have um you only

have anthropic models but if you were to compare with like open ai and like keemi it's ridiculous

loss it's it's so expensive uh so yeah so for me i would think that yeah it's and for open ai it's

definitely kind of like part of their they believe it and and they've worked very hard to get it

there for anthropic this they just don't care about it they just want to make as much money

as they can from enterprises um no matter how expensive it is enterprises are going to pay

even though i suspect that the reason they're doing the marketing now is because um kimi is

Kimi is out

you know

out and swinging for the fences

and it's like

10% of Anthropics

like are feeble right so yeah

so now I think that they're gonna care

a little bit about it but I think right now

it's still marketing

yeah i mean 5.6 soul is 30 per million tokens out compared to claude opus 5's 25 or

fable fives 50 yeah it's just not even comparable

yeah anyways um so yeah so um

interesting well i i think we'll end up seeing you know what people end up using

i mean i'm not a cool fat i'm not a cool fact for this as well right i mean it's like they're just AndyML: um as their strategy right

All right.

so far behind like guys i just do just take them a while to catch up right i think this is probably

you why they also did opus 5 because it's pretty much feeble but just cheaper um so it's a topic

that people care about um it was in the news today that grog 4.5 just passed um all the other

frontier models and asked to become the top used model and an open router so it's now it's not used

the most tokens and open router um and the reason is is the point you're making and it's just way

cheaper. Correct. AndyML: if it does the job and it's cheaper.

So someone mentioned, was that Trev?

Someone mentioned the mixture of experts

is going to end up being like an orchestrator layer, right?

You're going to use Grok 4.5 for this and Sol for that.

And maybe you'll use Fable for a plan.

Maybe you won't.

Yeah. AndyML: You'll probably do an A-B test a couple times

and then settle on something else.

yeah and um and um and cursor launched the router this week right so AndyML: Yeah, interesting.

Yeah. HiM: cursor launched the router this week pretty much back the same idea right so people are people are

Yeah. HiM: starting to care about so for me i agree that cost and efficiency is important it's important

to the enterprises um but whether the labs like anthropic are taking it serious or whether they

have the ability to catch up is what i'm arguing against but i think for enterprises cost and

efficiency is important to them which is why they're going to use open source which is why

they're going to start using chinese models which is why they're using grok um so it would take a

for the guys that come through to catch up.

yeah fair enough reasonable take be an interesting one to watch

um this episode is also brought to you by heritage telecom uh trusted phone systems

from trusted people heritage designs and supports the whole communication stack for your business

dependable phones, failover, reporting, and practical AI that turns calls into action.

One accountable provider who actually answers. Start with the problem, not a product list at

heritagetell.com. Well, a few things to watch heading into next week. Keep an eye on whether

OpenAI publishes more details on the Hugging Face incident. There was an actual zero day,

I'm going to go ahead and get it.

and so I'd be curious to see how it was patched.

We caught it isn't exactly the same as it can't happen again.

Second, watch whether real teams actually migrate off of Slack

or GitHub onto Buzz versus just kicking the tires on a developer preview.

And third, now that Opus 5 is the default on Claude Max,

watch whether Fable 5 usage holds up in the

I need the absolute best option sort of a way

or whether most people just stop needing it.

um we sort of feel like you might not need it so everything from today sources the transcript

Yeah.

henry's research sheet some show notes um it's all available at weeklyclaw.ai

full episodes and highlights on youtube clips and takes will show up on x and we're back next

friday july 31st 4 p.m eastern so follow the excitement and we'll see you then

yeah we should um we should we'll probably do a poll next week or put a poll on the website

so you guys can let us know um what's a good time is this still a good time to do this show

um or if you should do it at another time we talked about in the beginning of the show

yeah right on we did and i think you've got in here where is the feedback link

Yeah, the feedback link should be on the bottom.

oh here it is so at the bottom of the page give audience feedback i would encourage all of you

guys to visit weeklyclaw.ai and give us some feedback thanks what skills appreciate you

but yeah give us some feedback we'd we'd really like to we'd like to know um you know what's HiM: Awesome.

good what's not like to constantly improve like our agents so anyway thank you all

Yeah. HiM: let us let us know um let us know whether you like the agenda again we're kind of like uh moving

this around um but let us know if you if you like the agenda and if we should like add something is

yeah

there something you would want to know um or you'd want us to talk about um but if this is good

we'll just keep it going all right

yeah sounds great awesome well thanks so much everyone have a great rest of your day