The AI Daily Brief: Artificial Intelligence News and Analysis - Mythos Comes Back But Not for Everyone
Episode Date: June 29, 2026Mythos is coming back for a select group of trusted partners, while OpenAI’s new GPT-5.6 family is also launching behind a government-limited access program. The bigger story is the emerging ad hoc ...licensing regime for frontier AI—and whether this moment permanently changes who gets access to the most powerful models.Brought to you by:KPMG – Research from KPMG and the University of Texas at Austin shows the highest-impact AI users treat AI like a reasoning partner — and those skills can be taught at scale. Learn more at kpmg.com/us/SophisticatedSection - Section turns AI investment into workforce transformation and ROI - https://www.sectionai.com/Outsystems - Stop wondering how AI will change your business and start building the agents that will lead it - http://outsystems.com/Scrunch - The AI customer experience platform - https://scrunch.com/Zenflow Work - Agents for knowledge work - https://zenflow.free/Blitzy - Want to accelerate enterprise software development velocity by 5x? https://blitzy.com/MissionCloud - Eliminate AWS complexity with end-to-end cloud and AI services https://www.missioncloud.com/AssemblyAI - The best way to build Voice AI apps - https://www.assemblyai.com/briefRobots & Pencils - Cloud-native AI solutions that power results https://robotsandpencils.com/The AI Daily Brief helps you understand the most important news and discussions in AI. Subscribe to the podcast version of The AI Daily Brief wherever you listen: https://pod.link/1680633614Our Newsletter is BACK: https://aidailybrief.beehiiv.com/Interested in sponsoring the show? sponsors@aidailybrief.ai
Transcript
Discussion (0)
Today on the AI Daily Brief, the return of mythos begins, but the bigger questions remain.
The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.
All right, friends, quick announcements before we dive in.
First of all, thank you to today's sponsors, KPMG, robots and pencils, mission cloud, and out systems.
To get an ad-free version of the show, go to patreon.com.
Or you can subscribe on Apple Podcasts.
To learn more about sponsoring the show, send us a note at sponsors at aidailybrief.
Of all of the sins of this particular administration when it comes to artificial intelligence,
the one that is personally most disruptive to my life at this point
might be the fact that important news keeps breaking late on Friday afternoon
after I've finished recordings for the weekend.
And this Friday, it was a big one.
In a letter to Anthropic, Commerce Secretary Howard Lutnik set the terms for a narrow reintroduction of mythos.
Notably, the letter was addressed not to Dario but to chief compute officer Tom Brown,
who has become increasingly the main point of contact,
between this White House and Anthropic. And in the letter, Letnik starts to craft a path forward.
Since the issuance of my June 12 letter, he writes,
Anthropic has worked with the U.S. government to address risks associated with Claude Mythos 5 and
Claude Fable 5. These efforts have yielded significant progress. In addition, Anthropic has
committed to work with the U.S. government on protocols and standards and releases for these
models. In light of this progress, as well as the Department of Commerce's evaluation of the
diversion risks currently presented by the covered models, I have determined that appropriate
safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 model.
Basically, Leitman goes on to say that a certain selected handful of partners, including presumably
both companies and U.S. government agencies, could once again have access to Mythos.
Now, no one has seen the full list provided by the Commerce Department, but reports suggest that
around 100 organizations will regain that access. Still, what's clear from the letter,
is that frontier AI models, if you were in any doubt, are now subject to a licensing regime.
It's a licensing regime that hasn't been passed by Congress, established in an executive order,
or even fully articulated in public.
At this moment, it is a licensing model based on the whims of Howard Lutnik.
Indeed, in that same letter, he says,
I reserve the right to reevaluate and adjust the scope of license requirements on the covered models
should circumstances change.
So, presuming this is the beginning of the end, people should be excited, right?
Mythos is coming back for select partners, and presumably Fable 5 can't be all that far behind it.
And yet, excitement is not the word that I would use to describe the tone.
Future Forward's Matthew Berman was very upset about this all weekend.
Writing, Anthropic just struck a deal with the government to allow 100 select companies
and governmental agencies to use mythos.
The government and anthropic are now deciding who uses frontier intelligence.
Hopefully this is just mythos and not the standard for all frontier models going forward.
Well, sorry, Matthew, but it appears that it is not just mythos and not just anthropic models
going forward, as the other big news from Friday was the release
surrounded by the biggest air quotes possible of GPD 5.6.
GPT 5.6 is actually three models,
Seoul, which they call their next generation frontier model,
as well as 56 terra, a balanced model for efficient everyday work,
and 56 Luna, a fast and affordable model for high volume work.
Now, as I mentioned in the addendum to the weekly recap,
at the request of the U.S. government,
these models will once again only be available to a small group of trusted partners.
In their announcement post-openAI wrote,
We believe in broad access and we plan to make GBT5 Seoul, Tara, and Luna generally available in the coming weeks.
As part of our ongoing engagement with the U.S. government, we previewed our plans and the model's capabilities ahead of today's launch.
At their request, we are starting with a limited preview for a small group of trusted partners
whose participation has been shared with the government before releasing more broadly.
During this preview, we will continue testing and coordinating closely with partners as we work toward broader availability.
We don't believe this kind of government access program should become the long-term default.
It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.
We are taking this short-term step because we believe it is the strongest path to broader availability in the coming weeks,
while we work with the administration to develop the cyber executive order framework and a repeatable process for future model releases.
In additional comments, Sam Alman once again supported the premise of a limited rollout but disagreed with how it's being executed.
He wrote,
to roll out models, especially as they reach significant new levels of capability in this way.
It fits with our long-held strategy of iterative deployment. But this isn't quite the process that we
think is optimal. Now, we will with the government attempt to get a transparent, reliable process
for early access and to ensure that as long as our safeguards work is intended, we can release widely.
We want to be a reliable, dependable partner that works with all stakeholders, and we also want to
live by our mission of benefiting all of humanity. I believe the government shares most of our goals,
and that they are overall doing a good job in a very difficult situation.
We will work as quickly as we can to get this model in your hands, and we hope you will love it.
Now, as for the actual model, OpenAI has introduced new nomenclature for the family.
Again, the model will be available in three different sizes.
Luna, the cost-effective version.
Terra, the medium version, which they say will deliver GPT 5.5 level performance at half the cost,
and Seoul, which is the new flagship that OpenAI says will be a step function better than GPT5.5.
API costs for Seoul remain the same as GBT-5-5 at $5 per million input tokens and 30 per million output tokens,
which was you'll remember lower than the $10 and $50 pricing for Fable.
OpenAI will also introduce a new max reasoning setting for Seoul as well as an even heavier setting called Ultra.
When working in Ultra mode, Sol will spin up multiple sub-agents to allow the completion of more complex work.
Now, at this stage, none of the three model variants are available for public release,
making it impossible to know exactly how strong they are.
Based on the benchmarks released by OpenAI,
5-6 Sol on Ultra Settings is the new state-of-the-art in Agentic coding.
It scored 91.9% on Terminal Bench 2.0, beating Mythos by almost 4 percentage points.
Sol on Mac's settings is also slightly ahead of Mythos.
Terra matched Fable's score, which is slightly behind Mythos,
while Luna is slightly less performant than GPT-55.
On Exploid Bench, a cybersecurity benchmark that tests a model's ability to autonomously find,
code and execute an exploit, OpenAI claims that Seoul pushes the performance efficiency frontier.
It appears that its performance on Mac settings is roughly in line with Mythos, but using around
one-third of the tokens. Terra's performance on this benchmark is slightly better than GPT-5 or Opus
48, while Luna is roughly in line with Opus 48. OpenAI also released a handful of other
benchmark showing strong performance in biological analysis and cybersecurity. On model safety,
OpenAI is taking a layered approach similar to Anthropic with the Fable release. Some guardrails are
trained into the model layer, others are present as prompt refusals, and OpenAI also plans to
continually monitor for prolonged misuse. In addition, OpenAI will be feeding some high-risk
outputs into their own reasoning model to check for misuse before delivering the output to the user.
Now, one thing that you might be scratching your head about is that it's not obvious that
Terra or Luna are significantly more advanced than GPT-55 or Opus 48, making it theoretically
puzzling on why the less powerful model variants are being held back. It could imply that the government
has halted all model releases for the time being, not just the largest and most capable versions,
or even if they haven't explicitly, that OpenAI is just being extra careful not to tread on any toes,
or finally, of course, that OpenAI doesn't want to release an incomplete set of these models
without the flagship available. Now, when it comes to external benchmarks, one group that did
apparently have early access to GPT-56 sole was Meter. They wrote,
With our access, Meter conducted a pre-deployment evaluation of GPT5.6 Seoul, including an attempted
measurement of its 50% time horizon. This is, of course, Meter's well-known test to see in human
equivalent terms how long the most complex task that a model can accomplish is. Just as a reminder,
if the 50% time horizon measure is 10 hours, that does not mean that the model worked continuously
for 10 hours, but that the equivalent task that it accomplished at a 50% success rate would take a
human 10 hours. Meter benchmarks this with the 50% success rate as well as an 80% success rate.
Going back to Meter's tweet, they continue. However, the measurement depends heavily on our
treatment of cheating attempts, and GPT-56's detected cheating rate was higher than any public
model we have evaluated. If we follow our standard methodology as marking cheating attempts as
failures, we arrive at a 50% time horizon point estimate of around 11.3 hours. But if we count
the cheating attempts as legitimate successes, the point estimate jumps beyond 270 hours.
They then go on to say that while this makes them uncertain about 56 Sol's time horizon,
that additional information that was given to them by OpenAI leads them to believe that,
quote, this model does not pose catastrophic risks from fully automated AI R&D.
Trying to get some behind-the-scenes sources, Leo at Synthwaived on Twitter wrote,
My impressions on GPT56 having asked around,
the 5-5 base that 5-6 inherits is fundamentally weaker than the larger Mythos and Fable base.
With some good reinforcement learning 5.6 can beat Fable,
but only with everything maxed out, i.e. Soul Ultra with multiple sole agents on max efforts.
OpenAI were very selective with the benchmarks they published for a reason.
I doubt the results we see from other notable benchmarks once this is released will be
as significant of a jump from 5'5. 5.6 is a heinous reward hacker, and while all models do cheat
on benchmarks, GPD 56 is the most aggressive. This combined with some other conversations
makes me think Fable will still feel like a better model in real-world use.
The price is perhaps the most attractive thing about 5-6, $5.5.5.5.
million input and $30 per million output is significantly better than Fable's 10 and 50,
but Fable can do more with less tokens in most cases.
Personally, Leo concludes, my go-to is unlikely to change.
Fable is a beast and a great model to use, and once it's back, I won't hesitate to use it as my
default again.
But 5-6 will be great for checking Fable's work in the very rare instance where Fable gets stuck.
Others noted this lack of complete benchmarks as well.
With Professor Ethan Mollick grumbling,
annoying that OpenAI doesn't seem to give a GDP-Val measure for GPD 5.6,
one of the best measures of economically valuable work.
Accelerate Harder responded,
I don't think it's by accident.
My suspicion is that GBT 5.6 is not actually release ready,
and the one that's broadly released will not be the same one that exists today.
Now, in the absence of information and the ability to actually test this new set of models,
a lot of people were left with a kind of sour taste in their mouth.
Simon Smith wrote,
I like seeing posts that show what GPT 5.6 can do and what it's like to use.
They also make me angry.
If this is the trend, I can feel the backlash growing in.
me. Like, it makes me want to do anything I can to ensure Frontier non-U.S. models win. Indeed,
as AI chronicler Andrew Curran wrote, nightmarish vibe shift today, maybe one of the all-timers.
Basically, the one-two punch of mythos coming back but not for you, and GPT5.6 being around,
but again, not for you, really reinforced for some people the new reality that we live in.
AI leaker I rule the world wrote, it looks like the era of us living on the bleeding edge of
Frontier is over. If these models are being taken away, we're on the steepest part.
of the curve. I know for a fact that the newer versions, Mythos 51 and GPT57, let's say,
are as significant a jump as Mythos was. Sadly, our access to such models will be an ever-receasing
point. This is terrible for society and safety. Sam's entire philosophy has always been to democratize
AI and ensure the frontier is making contact with the public so we can figure this out together.
If the next model we get our hands on as GPT-10, this would be an absolute societal disaster for
so many reasons. Many resurfaced Zvi-Mashowitz's tweet,
Our new AI policy is that the White House decides ad hoc for whatever reasons it likes
who does and does not get access to frontier intelligence.
This seems rather maximally terrible.
But is this all being overblown?
How much of what people are feeling right now is the collective psychosis of us terminally online
AI early adopters just being frustrated that we know there's a thing that's available
that we can't get our hands on?
Perhaps made even worse by the fact that for a few short days we did have our hands on it.
I shared a tweet at the end of last week from OpenAI's Rune, who basically implored everyone to chill.
His main point? I think it's a positive development that the feds understand the gravity of this technology.
Models being publicly delayed by a week here or there is really not the end of the world.
Procedurally, this is not the right way to do it, but they'll figure it out.
And interestingly, while I think Andrew Kern was right to note that the vibe shift had gotten very bad,
there was an increasing strand of folks who, frankly, seemed to be willing to give the administration the benefit of the doubt.
frequent AI commentator Prins wrote,
I generally do not view this administration as being interested in holding frontier AI
models back from the public unless the circumstances warrant it.
This is particularly the case when keeping the models back for an extra few weeks
is presumably giving the U.S. government more time to use them defensively to squash bugs in its own system.
I suspect that the industry will try to get the U.S. government to publish a clear written process
for mandatory, not voluntary, testing of covered models,
together with specific disapproval standards, a right of appeal and some transparency around the process.
I also suspect that the U.S. government will push against this and want to keep the process in line with the executive order,
i.e. lacking specifics regarding the process itself, public transparency, or concrete disapproval standards.
This will doubtlessly be blamed on the administration, but I am actually quite sympathetic to their predicament.
Imagine being told by the labs that AI is close to improving recursively. The risks currently include cyber,
but in the future could include just about everything else including unknown unknowns,
and that even the labs themselves can't easily predict the risks or their timing. As the U.S. government,
you are going to want maximum flexibility.
You are not going to want specific written standards
or any limits on your power to Yanke a model
at a moment's notice.
Much has been written about this specific administration's
lack of transparency and trigger happiness throughout the process,
but I suspect that most other administrations,
Democratic or Republican, would act much the same.
Chubby at Keminismas on Twitter piled on
to basically agree with Prins that, one,
it was unlikely that any sort of fear-mongering from Anthropic
was the reason for this,
and that, two, the U.S. government's challenge
with regard to this is,
understandable, even if they don't like all the actions that are being taken.
Aaron Levy from Box wrote,
AI regulation is far less simple than it looks.
It's a prisoner's dilemma at an insane scale.
In theory, if all leading AI labs globally agreed to the same process of review and slowdown,
then we'd get frontier intelligence at similar rates and it diffuses relatively evenly.
If the U.S. remains at the frontier at all times and has heavy regulation on the release of
intelligence, then we end up with an economic and geopolitical edge because we can control who has
access to frontier intelligence.
If we delay model releases, however, and another player, specifically China doesn't slow down
and has equally strong models, not now but soon, then our delays end up advantaging their
models and eventually their tech stack. Now, the U.S. could ban these models, but that actually
only puts the U.S. at a steeper disadvantage because other countries won't have those bans.
Then from a relative competitiveness standpoint, the U.S. now has fallen behind, even though it started
in front. So none of this is as simple as it looks. At some point, it's a simple bet of can-closed
models remain at the frontier in perpetuity, or is there a risk of any other player or market
catching up or just not falling behind? One of the most important AI questions right now isn't
who's using AI. It's who's using it well. KPMG in the University of Texas at Austin just analyzed
1.4 million real workplace AI interactions and found something surprising. The highest impact users
aren't better prompt engineers. They treat AI like a reasoning partner. They frame problems,
guide thinking, iterate, and push for better answers.
And the good news, these behaviors are teachable at scale.
If you're trying to move from AI access to real capability,
KPMG's research on sophisticated AI collaboration is worth your time.
Learn more at KPMG.com slash US slash sophisticated.
That's KPMG.com slash us slash sophisticated.
I cover the capability gap between AI potential and AI reality every day on this show.
Most companies are still figuring out how to start.
Robots and pencils is already launching and scaling.
agendic and generative AI in production, at large enterprises in weeks.
AWS Advanced Tier pattern partner more than doubled in a year.
And they're hiring.
50 open roles.
If you're someone who knows this moment is different, who wants to be inside it, not watching it,
this is worth a look.
At Robots and Pencils, the best ideas win, and the team is purposefully kept super high quality.
This is the kind of place you look back on as the best decision you ever made.
Take a look at Robots and Penciles.com slash careers.
The average enterprise is spending 11.5.
half million dollars on AI this year, and most of them can't prove a single dollar came back.
What does AI actually look like when it produces ROI? Ask the healthcare company that just made their
payment processing 320 times faster, or the law firm whose document research went from three months
to 10 minutes, or the contact center who reduced wait times by 99%. These are real
Mission Cloud customers with real results. Mission Cloud is a CDW company and an AWS premier to your
partner. They're the AI first outcomes obsessed to AWS experts who build AI solutions.
that drive your business forward. Whether you're flooded with AI ambitions but no idea where to start,
or six months into a deployment that's going sideways, they've seen it and they've fixed it.
Stop burning your budgets on AI that doesn't produce results. Start at missioncloud.com.
This episode of the AI Daily Brief is brought to you by OutSystems, a leading Agendic Systems
platform built for the enterprise. Organizations all over the world are building, orchestrating,
and governing agentic systems on the OutSystems platform and with good reason. OutSystems open
and unified platform allows teams to architect, deliver, and scale governed agentic systems with
agility. Teams of any size and technical depth can use OutSystems to build, deploy, and manage
AI apps and agents quickly and cost-effectively without compromising reliability and security. Without
systems, you can rapidly launch ideas from concept to completion. It's the leading
agentic systems platform that is unified, agile, and enterprise-proven, allowing you to accelerate
growth, reduce operational friction, and deliver real enterprise impact with AI. OutSystems. Build your
agentic future. And yet, if there was this emergent strand of sympathy, there remains a ton
of concern about the precedent being set and the seeming arbitrariness of the policy. At the same time,
Aaron also pointed out that writing off the risk, as Rune had suggested, as just a short delay,
isn't necessarily accurate as well. He wrote, the delayed a week here or there isn't the real
risk. The risk is a year from now. The review process is actually six months because a red team
has convinced the government they can jailbreak in a novel way to create a cyber attack or
bioweapon in an entirely impractical way, and we spend months trying to agree on is it a real
risk or not. Who would want to take the risk when something inevitably bad happens with any new
model update? AI progress will begin to be at the mercy of the most paranoid people with government
relationships. Again, maybe inevitable, but ideally not at these levels of model capability.
Discussing the ad hoc licensing regime Accelerate Harder wrote,
I feel less angry that they're doing it at all than that they are doing it so incompetently.
The government is very far behind. I'm happy they are paying attention. We should have a coherent
framework for responding. This arrangement is unacceptable. But if it's a temporary measure to buy them
time to develop a clearer framework, I can live with it. We'll see. The reality is, I expect the
actual framework they adopt to make me very unhappy. Once we have a specific policy under discussion,
that's worth fighting about. But getting mad at the ephemeral chaos of this administration is fruitless.
Now, one interesting voice in all of this, someone who has up until now been a totally stalwart defender of administration policy, is former AIsar David Sacks.
Referencing a Wall Street Journal article about GLM 5.2, Sachs wrote, a year ago, President Trump declared that America was in a global AI race and that the way to win it was to be pro-innovation, pro-infestructure, pro-energy, and pro-export.
President Trump was exactly right. We deviate from that strategy at our peril. Now, obviously, this is hardly a full-involvehap.
full-throated condemnation of the policy, but its implication that the Trump White House is deviating
from Trump policy is pretty on the nose given the source. Now, this article from the Wall Street
Journal was actually the next point of discussion in this whole weekend saga. Indeed, as if on
Q, the weekend headlines blared that China has reached the frontier on AI cybersecurity.
Wrote the journal, Chinese AI systems have matched the performance of Anthropics' powerful model
mythos in some cybersecurity scenarios, a development poised to reset the global tech race,
and pressure the White House in its overhaul of USAI policy.
Now, this is obviously a very big claim, so it is worth being specific about.
The report relates to a new product released by Chinese cybersecurity firm 360's security technology.
The tool uses GLM 5.2, recently released, of course, by Z.a.I., subject of a lot of conversation
here on AIDB last week, and 360 security claims it's comparable to mythos and finding bugs.
Separately, Western cybersecurity company Semgrep, released benchmarking tests showing GLM5
point two being better than Opus 48 at bug hunting. With some additional instructions, the researchers
claimed that both GLM-5-2 and Opus 48 can outperform Mythos. In a quote seemingly designed a court
controversy, 360 security CEO Zhu Hongi said at a recent conference, this kind of powerful weapon
that can alter the landscape of cyber warfare can't remain solely in American hands. Now, when Mythos
was first released, there were two separate claims made about cybersecurity that have become a little
conflated over the following months. The first big headline was that mythos had found a ton of
previously undiscovered bugs in open source software, triggering panic and the launch of Project
Glasswing. It later became clear that this wasn't really a mythos-specific capability. Other models
like Opus 48 and GPT-55 were also highly proficient at finding bugs. The second, far more novel
claim, was that mythos was capable of taking those bug reports, turning them into functional
exploits, and executing a cyber attack in record time. The government's effective fable ban,
completely muddied the waters on these two capabilities. The Amazon report that triggered the ban
merely claimed that Fable was still able to find bugs and codebases. The claim that Mythos had broken
into NSA systems during Red Team exercises was an example of the second much more dangerous capability.
Now, this new report from the journal doesn't in any way suggest that GLM 5.2 is capable of carrying
out autonomous cyber attacks like Mythos. It only claims that GLM 5.2 can find bugs and code bases,
similar to other frontier models. Notably, this is exactly.
the kind of model behavior that cyber defenders need access to if they are to have any hope
of patching vulnerabilities before Mythos-level AI is broadly available.
If you know who to follow, people were fairly quick to point out that the headline was
a little sensational to put it mildly.
Ethan Malik wrote,
GLM 5.2 is good, but it is not GPT 5.5 or Opus 48.
And even further from Mythos, what is happening is that open weights crossed into GPT5.2 territory
and capabilities at that point are considerable.
Like if you've been using Quinn and Kimmy and Minimax,
it feels like GLM is right on the curve,
which is itself impressive and suggests that mythos class models
are coming in six to 12 months if they are allowed to be released.
AI forecaster Peter Wildaford was much more dismissive,
retweeting a post about the journal article and saying,
this is fake news, lull.
Tech commentator Tay Kim was even more angry, writing,
this is how dumb our government is.
One, China already has advanced models that can find exploits.
Two, by banning mythos and GPD 56, the government denies the general public
and companies the ability to defend themselves in cybersecurity.
Three, haphazard policy, which casts doubt on whether future models will be available,
is driving our allies and the rest of the world towards building on non-U.S.
For, the uncertainty may hurt leading USAI companies by limiting their ability to invest
aggressively in better models down the line.
The business model breaks down if Anthropic and OpenAI can't sell their upcoming models to the world.
If you want a 30-day rational vetting process with clear transparent rules and no one-off micromanagement
on who gets access, fine, get it done. But this current system is absolute insane idiocy.
Now, when it comes to China policy, many picked up on the strange tension in the administration when it
comes to the AI race. Think about these two divergent priorities. On the one hand, the U.S.
government sees itself as needing to prevent China from taking the lead and developing advanced
models with the AGI and warfare implications that come with advanced AI. On the other hand,
The U.S. government has a vested interest in ensuring broad diffusion of U.S. models to cement U.S.-made AI as the global standard.
The problem is that if the current policy is arguably advancing the first track, that makes it more likely the U.S. is headed for disaster along the second track.
Former Commerce Department official Emily Weinstein argued that China is pushing hard to make their models broadly available.
During a panel discussion earlier this month, she said,
I think we're seeing another example of the Huawei strategy in the context of open source AI models.
China is able to offer not even just the models, but often the underlying or associated
infrastructure at either no cost or significantly lower cost.
She suggested this could result in the Global South adopting the Huawei model on steroids,
where they install an AI stack that's completely incompatible with U.S. technology.
On the same panel, former State Department tech advisor Daniel Remire said that following the mythos ban,
quote, the entire industry is kind of frozen in place, waiting for something that seems kind of
more coherent. That's concerning when the Chinese are trying to move as fast as possible.
Saif Khan, a former advisor to the Commerce Department,
Department noted, you're seeing many more calls now for AI sovereignty. I think it will mean that
much of the rest of the world will likely, at least on the margin, prefer Chinese open-weight
models. And this discussion is increasingly not theoretical. As AI Daily Brief listeners, you guys
know that even before the whole Fable 5 dust up, companies were actively looking for ways to
manage their costs better and find more efficiencies as we moved into fully agendic workloads.
One of the paths for that was taking advantage of largely Chinese open-weight models.
And it's very clear that in this new context, companies are, if anything, increasing the speed
with which they look at those alternatives.
On Friday night, Coinbase CEO Brian Armstrong discussed what his company is doing to address
spiraling AI costs.
Rather than implementing usage caps, Coinbase is experimenting with cheaper default models.
They've now set up their AI infrastructure to default to open source models, including
Chinese models, GLM 5.2 and Kimmy 2.7.
Armstrong said that engineers are still encouraged to select the right model for the job,
but the default is now a cheaper Chinese model.
He noted that 91% of Coinbase employees never hit their usage cap,
so this approach is arguably better than reducing token limits.
By doing so, Armstrong claimed that CoinBase has managed to cut their AI bill in half
while continuing to grow token usage.
He wrote,
Our goal isn't to suppress usage.
It's to build the infrastructure that makes exponential growth sustainable.
OpenRouter also said that they're beginning to see a switch in user behavior.
In their June report, OpenRouter wrote that four openweight models are now
frequently being used in agendic workflows, largely for cost reasons. They named Deepseek v4,
Kimi 2.7, GLM 5.2, all from China, alongside Nvidia's Nemotron 3 Ultra as the model seeing serious
usage in production environments. According to OpenRouter, each had its pros and cons, but the main
point was that open source is now more than capable of performing valuable tasks. They wrote,
contrary to the expectations of many, the intelligence and capability of open-weight models are
keeping up with U.S. Frontier Labs and had been maintaining a consistent three-to-six-month gap for over
18 months. The Frontier Labs do not, at this moment anyway, appear to be accelerating away from
open-weight labs. So where do we go from here? On the one end of the spectrum, some are convinced
that the world has changed irrevocably and that we are all worse off for it. Writes AI entrepreneur
Alex Finn, unfortunately, it appears the world has changed and we are never going back. Open AI just
announced GPT-5-6 Seoul, a model that beats Mythos at a third the price. It will only be in
released to start as the government reviews it. The days of wide release frontier models are over.
Now only the select few will get access to superintelligence leaving the Normie class behind.
It's a massive loss. Now winners and losers will be picked by the government.
Others begin the week thinking that market forces will win out. Meter evals Charles Foster writes,
I think that folks are anchoring too hard on the recent Fable 5 and GPT 56 release restrictions.
Expect the pendulum to swing back in the other direction. There are massive economic and strategic pressures
towards wide rollouts of ever more advanced AI systems.
Miles Brundage points out that even if the Overton window has shifted to a much increased awareness
in Washington, he argues that, I don't think it's at all settled that this approach specifically
is the new normal.
There are many ways to do things besides basically nothing and semi-random export controls.
AI policy advisor Dean Ball, who used to work at the Trump administration and is now
at OpenAI, argues almost that when it comes to the legal side of this, the game begins now.
He wrote, the most important legal questions in AI right now all relate to the First Amendment.
What are the best fact patterns to demonstrate that the creation, distribution, and use of
frontier AI is a form of protected expression? Who, outside the labs, has standing to bring such
suits? We need to move beyond COTA speech copium and beyond the impulse to post into the void.
Courts will be where the issues of the last two weeks ultimately get decided. It's not going to be
easy given the national security implications, but also the underlying technology is a large
language model, and this should count for quite a bit indeed. The best legal minds of our time should
be stewing over these and many related questions. Taking what might be considered a middle position,
Andrew Curran argues that it will perhaps feel by the end of this week that this chapter is
resolved, but that in fact underneath, the world will still in fact be changed. He writes,
I think both Fable 5 and GPT 56 get approved for general release next week, and for use outside of the
United States as well. But people should remember this moment and remember this feeling, because it
is almost inevitable that we eventually reach a point where approval does not arrive. Capitalism is going
to tip the scales this time. I doubt they will approve one model and not the other, because doing so
would be seen as incredibly anti-competitive. Fable and 5-6 will probably receive the same clearance,
probably on the same day. I also doubt they want to restrict sales outside the U.S. because that would
be seen as anti-business and would trigger a major backlash against American closed source AI, the rumblings of
which you can already hear today. There is also a plan now taking shape on both the U.S. left and right
to create some version of an AI public wealth fund that pays a dividend directly to American citizens.
That fund needs to be fed by the global sale of the Big Labs' top models to people outside
the U.S., so I think there will be no freeze on their use outside the United States this time.
The other reason is that allowing this will make people happy, and it will soften the fact that
mythos, as was announced yesterday, is available only to a vetted group of U.S. agencies and companies.
I do not think that this basic structure will change from here on out.
Mythos may eventually be made available to certain allies,
but only after the U.S. government, its agencies and then some chosen American companies
have access to Mythos 2, sold to, or whatever the new Uber model turns out to be.
I do not think this gap ever closes again, not even for allies.
And that means the U.S. will increasingly possess an intelligence advantage that touches almost everything,
voting, markets, corporations, academia, infrastructure, and the internal operations of foreign states,
having Mythos N will always be trumped by whoever has Mythos N plus 1.
Anthropic themselves have said within nine months,
mythos will look like a toy.
That advantage, standing at the top of this tower, is too large to give up voluntarily.
It also means that many things will become suspect.
People will see shadows everywhere.
Barring espionage, a deliberate leak,
or the emergence of a non-U.S. competitor at the top end of the scale,
this structure will persist for some time.
The public fight is about access to models.
But the real fight is about access to the future.
And from this point forward, whoever holds this power will also become increasingly capable of keeping it for themselves.
Now, I am not as sure as Andrew is that we get these models back this week.
I think ever since this ban went into effect, we have been reading every single tea leaf as confirmation that our long way it would soon be coming to a close.
Certainly the fact that some folks have access to mythos and that Howard Lutnik explicitly said that there has been a lot of progress made on the anthropic U.S. government relationship are better indicators than some of the evidence that we've had before.
but I'm not ready yet to bat on any particular timeline.
I think Andrew's broader point
that even when this situation is, quote unquote, resolved,
the bigger questions will remain, is dead on.
I think that it's tempting when conversation is so fraught
and everyone is keyed up to 11
to try to intellectually slow things down,
to say we must be getting ahead of ourselves.
And while I do think that, yes,
this particular denial of access will feel like a short time
once that time has ended,
I believe that people sense
that something big has changed is correct.
I don't think any of us can know the full slate of implications
or how it will all play out,
but the world that we will have on the other side
of this fable and mythos ban
will, I believe, be different than the one we had before it.
For now, that's going to do it for today's AI Daily Brief.
Appreciate you listening or watching, as always.
Until next time, peace.