WEBVTT
00:00:01.600 --> 00:00:10.800
Normally, when OpenAI releases a new model inside of ChatGPT, it has the potential to instantly change how millions of people work.
00:00:11.199 --> 00:00:16.800
Think new capabilities, support for new inputs, and new artifacts as outputs.
00:00:17.039 --> 00:00:25.839
You can see why 1 billion people getting access to something as straightforward as a model update could actually be a huge deal.
00:00:26.000 --> 00:00:32.719
But with OpenAI's latest model update, GPT-5.6, we didn't really get any of that.
00:00:32.880 --> 00:00:39.840
That's because for the first time we got a GPT upgrade that was announced but not released.
00:00:40.079 --> 00:00:50.799
That's because ever since Anthropic escalated its feud with the federal government, Frontier Labs and the US have been under increased scrutiny to ensure its models are safe.
00:00:51.039 --> 00:01:01.679
So on Friday, when OpenAI unveiled GPT-5.6, all but like 99.9% of the world just got a blog post and a TVD.
00:01:01.920 --> 00:01:13.120
Yet even though only a select few trusted companies in the world can use GPT 5.6 right now, that doesn't mean you should sit idly by.
00:01:13.359 --> 00:01:15.359
Actually, I'd argue the opposite.
00:01:15.519 --> 00:01:25.519
Because with this new government AI permission slip era that we're all living in now, the smartest thing you can do is actually understand the new rules before everyone else does.
00:01:25.680 --> 00:01:33.439
Because the leaders who figure out how to operate in this world right now are the ones who will have a massive head start when the doors finally open for everyone.
00:01:33.680 --> 00:01:37.680
So today I'm going to show you exactly what the new GPT-5.
00:01:43.840 --> 00:01:46.239
So here is the big picture.
00:01:46.400 --> 00:01:51.599
Well, there's a powerful new model, but it's kind of locked up at launch.
00:01:51.680 --> 00:01:59.120
So OpenAI announced GPT 5.6 on Friday, but they blocked normal ChatGPT access to it.
00:01:59.200 --> 00:02:07.120
So normally when OpenAI releases a new model, well, everyone just goes to chatGPT.com and they start using it and making changes to the workflow.
00:02:07.280 --> 00:02:08.000
Not the case now.
00:02:08.080 --> 00:02:09.360
It's a limited release.
00:02:09.439 --> 00:02:14.960
Uh only a uh couple dozen uh companies in the world reportedly have access to it now.
00:02:15.120 --> 00:02:17.199
There's also a new naming system, which we're gonna get to.
00:02:17.280 --> 00:02:27.919
So it is called GPT-5.6 Soul, Terra, and Luna, and it packages frontier capabilities into what OpenAI is now calling durable product tiers.
00:02:28.080 --> 00:02:31.039
So OpenAI previewed the model to the US government first.
00:02:31.199 --> 00:02:35.439
We're gonna get into that, then limited the launch at the government's request.
00:02:35.680 --> 00:02:45.439
And this whole situation right now mirrors the harsher federal move against anthropic weeks earlier as AI now becomes controlled infrastructure.
00:02:45.599 --> 00:02:48.800
So on today's show, stick with me for the next 20-ish minutes.
00:02:48.960 --> 00:02:58.080
Here's what you're gonna learn you're gonna know what OpenAI actually built and why even its cheapest model now rates as dangerous in the new 5-6 series.
00:02:58.240 --> 00:03:06.479
You're gonna know the varied admission from OpenAI's own tests that show that this model kind of cheated and overstepped its own instructions.
00:03:06.639 --> 00:03:12.159
You're gonna know why the federal government gated access and how this echoes the in-propic shutdown weeks earlier.
00:03:12.319 --> 00:03:21.759
And I'm gonna give you the operator playbook for when the best model is well one you can't actually use and what you should do about it in that case.
00:03:21.919 --> 00:03:23.680
All right, I'm excited for today's show.
00:03:23.759 --> 00:03:24.479
Hope you are too.
00:03:24.560 --> 00:03:25.759
This is Everyday AI.
00:03:25.840 --> 00:03:26.159
Welcome.
00:03:26.240 --> 00:03:28.479
My name's Jordan, and we do this every day.
00:03:28.560 --> 00:03:30.560
That's the name, Everyday AI.
00:03:30.639 --> 00:03:43.439
So this is your daily live stream podcast and free daily newsletter helping everyday business leaders like you and me not just keep up with the AI developments, but I tell you what it actually means in plain technical, non-technical terms.
00:03:43.599 --> 00:03:45.599
I give you the non-BS version.
00:03:45.680 --> 00:03:48.240
You take that information to grow your company and your career.
00:03:48.400 --> 00:03:51.599
So it starts here with the unedited, unscripted live stream podcast.
00:03:51.680 --> 00:03:55.520
But make sure you go to our website at your everydayai.com.
00:03:55.599 --> 00:03:57.039
Sign up for the free daily newsletter.
00:03:57.120 --> 00:04:02.879
We're gonna be recapping today's show and everything else that you need to know in the world of AI.
00:04:03.039 --> 00:04:06.159
There's a ton of late breaking news over the past couple of hours.
00:04:06.240 --> 00:04:08.000
So make sure you check out today's newsletter.
00:04:08.159 --> 00:04:08.800
All right.
00:04:10.400 --> 00:04:10.960
Ooh.
00:04:11.280 --> 00:04:12.719
GPT 5.6.
00:04:13.360 --> 00:04:14.960
We don't have it, right?
00:04:15.199 --> 00:04:16.480
Very few people do.
00:04:16.720 --> 00:04:22.720
And I think aside from we're gonna go over, you know, kind of OpenAI's blog post a little bit.
00:04:22.879 --> 00:04:31.120
Uh, I'm gonna tell you some of my big takeaways from looking at the model card that they released that I don't think a lot of people read or look at, because not everyone's a dork like me.
00:04:31.199 --> 00:04:38.079
And I'm gonna give you some of the reaction as well uh from some of the very few people who did get access and actually talked about it.
00:04:38.240 --> 00:04:41.600
Uh, but here's the stats.
00:04:41.920 --> 00:04:45.439
So this is now a three-tier model family.
00:04:45.680 --> 00:04:54.079
So GPT 5.6 is now one generation split into three lasting tiers that OpenAI alluded that they're gonna keep naming models this.
00:04:54.160 --> 00:05:03.600
So we don't know if we're gonna get a 5.7 or a you know, GPT six next, uh, but it seems like they're gonna keep using these naming mechanisms.
00:05:03.680 --> 00:05:09.759
So Sol, Terra, and Luna in the easiest way, because yeah, we're having to learn a bunch of new names now.
00:05:09.920 --> 00:05:13.839
I would just map this to Opus, Sonnet, and Haiku.
00:05:14.319 --> 00:05:14.879
Uh, right.
00:05:14.959 --> 00:05:23.920
So Sol is the most powerful, uh, Terra is the middle, and then Luna would be the one that's fastest and cheapest, but not the smartest.
00:05:24.079 --> 00:05:30.720
So uh, in the same way that Anthropic has kind of had their three classes, now we're gonna have those three dedicated classes.
00:05:30.800 --> 00:05:40.800
And to make it even a little bit more confusing, it looks like there's multiple intelligence tiers within those classes, but we'll get to that uh later when we all actually have access to this.
00:05:40.959 --> 00:05:43.600
But soul is the main event, all right.
00:05:43.680 --> 00:05:46.079
And if you're Spanish, isn't that great?
00:05:46.240 --> 00:05:50.560
Uh, you know, soul, terra, luna, that's sun, earth, moon.
00:05:50.639 --> 00:05:52.879
So maybe that'll make it easier to remember, right?
00:05:53.040 --> 00:05:54.399
Uh, third rock from the sun.
00:05:54.720 --> 00:05:57.839
Soul, uh, that is the the big dog, so to speak.
00:05:57.920 --> 00:05:59.920
So that handles the hardest reasoning and coding.
00:06:00.000 --> 00:06:05.519
Uh, terra in the middle balances costs with performance, and then Luna can just run faster and cheaper.
00:06:05.680 --> 00:06:11.519
So Seoul is the flagship and its strength set in the entire situation in motion.
00:06:11.680 --> 00:06:16.000
So uh here's the rollout.
00:06:16.160 --> 00:06:24.319
So, right now, um the companies that do have it reportedly are just using it via the API and codex.
00:06:24.560 --> 00:06:36.399
So OpenAI hasn't said that when and if this rolls out in a couple of weeks because they said that they plan to make GPT-5.6 Sold Terra and Luna generally available in the coming weeks.
00:06:36.480 --> 00:06:41.680
They didn't say where, uh, so right now it's only available on the API and codex.
00:06:41.839 --> 00:06:45.279
So we're not sure if this will be released inside of Chad GPT.
00:06:45.360 --> 00:06:57.839
My assumption is they will, uh, but I'm guessing they're using it uh right now with the trusted testers only uh via the API and codex because it's a little bit easier uh to have those guardrails in place.
00:06:58.480 --> 00:07:00.879
So right now it's just a few.
00:07:00.959 --> 00:07:04.399
So reports have said it's like 20-something companies that have this.
00:07:04.480 --> 00:07:16.319
Uh, and I'm sure there's a lot of people at those companies, so don't know exactly how many users, but at least with the 1 billion weekly active users that OpenAI has, it's safe to say that 99.9% of us do not have access.
00:07:16.399 --> 00:07:25.759
And when if we do get access in two weeks, we're still not even sure if it will be inside of Chat GPT or if it will only be in the API and in codecs.
00:07:25.920 --> 00:07:34.319
So uh if nothing else, if you're a power Chat GPT user and you're like, Yeah, I want to get my hands on GPT 5.6, well, maybe start using codecs a little bit.
00:07:34.480 --> 00:07:37.600
So the good thing is pricing didn't change, which is great, right?
00:07:37.759 --> 00:07:43.199
Anytime you see a nice jump up in capabilities and pricing stays the same, that's sweet, right?
00:07:43.360 --> 00:07:45.600
Uh so this on the API side, right?
00:07:45.680 --> 00:07:55.759
Uh, this is just if you're paying uh by the token, it is still Seoul is still five dollars uh per million input and thirty dollars per million output, which is the same as GPT 5.5.
00:07:56.079 --> 00:08:11.600
Uh and then Terra, about half that cost with 250 in per million, 15 out per million, and then Luna, very affordable, at one dollar uh per million input and six uh six dollars uh per million output, uh which is interesting, right?
00:08:11.680 --> 00:08:20.959
Because we talked about uh, like as an example, GPT uh GPT 3.5 flash being a much more expensive model than previously.
00:08:21.120 --> 00:08:25.680
And I know that Google had historically been cleaning up on the Flash series.
00:08:25.759 --> 00:08:31.680
So we'll see maybe here the uh story that a lot of people might not be talking about.
00:08:31.839 --> 00:08:34.559
Uh yeah, wait and we'll check back in.
00:08:34.639 --> 00:08:41.279
But I don't know, I think maybe a sleeper here is Luna being used by developers uh for simple things.
00:08:41.360 --> 00:08:54.720
I think that you know, one thing that we're gonna talk about as we talked uh or as we alluded to uh on our recent start here uh series episode, um, where we talked about going from token maxing uh to token efficiency.
00:08:54.879 --> 00:08:56.960
So go check out episode 789.
00:08:57.200 --> 00:09:01.759
But I think it's gonna become very common in the third and fourth quarter in the enterprise first.
00:09:02.080 --> 00:09:07.600
I think everyone else is gonna catch up to this in early 2027, but essentially using a mixture of models.
00:09:07.759 --> 00:09:08.240
Weird.
00:09:08.399 --> 00:09:18.799
Something I said two years ago in my bold predictions, uh, because it just doesn't gonna make sense, especially if you're paying per token on the API side, like so many enterprise companies are.
00:09:18.960 --> 00:09:38.080
It doesn't make sense anymore to be paying the same uh you know price per token for something that is just gonna help you rewrite an email versus something that's gonna go do a uh 30-hour autonomous project that would take a team refactoring, a code base, you know, two weeks of human hours, right?
00:09:38.159 --> 00:09:47.360
It doesn't make sense to use the same, you know, uh$5 input,$30 model as it would for something that's you know 20% of the cost.
00:09:47.679 --> 00:09:50.799
Uh so that's I would just say, you know, it's not out yet.
00:09:50.879 --> 00:09:52.480
We'll see, benchmarks, all that.
00:09:52.720 --> 00:09:53.440
But I don't know.
00:09:53.600 --> 00:10:03.519
To me, with Google uh increasing the price of their flash model, uh, anthropics uh models not getting cheaper anytime sooner, keep an eye on Luna, actually.
00:10:03.759 --> 00:10:05.519
So let's keep going.
00:10:05.679 --> 00:10:07.440
This is from OpenAI's blog post.
00:10:07.519 --> 00:10:12.399
I'm just gonna read from their words and then give you some other people's words and my words as well.
00:10:12.639 --> 00:10:24.639
So they said we're beginning a limited preview of the GPT-5.6 series, Seoul, our flagship model, Terra, a balanced model for everyday work, and Luna, a fast and affordable model.
00:10:24.879 --> 00:10:31.200
Terra has competitive performance to GPT 5.5 while being two times cheaper.
00:10:31.360 --> 00:10:35.279
And Luna brings strong capabilities at our lowest cost.
00:10:35.519 --> 00:10:40.720
GPT 5.6 Soul launches with our most robust safety stack to date.
00:10:40.879 --> 00:10:53.600
We strengthen protections for higher risk activity, sensitive cyber requests, and repeated misuse and spent multiple weeks finding weaknesses, pressure testing our system, and hardening it against real-world attacks.
00:10:53.759 --> 00:11:04.799
All right, and then they go on to say, we believe in broad access and we plan to make GVT56 Soul, Terra, and Luna generally available in the coming weeks.
00:11:04.879 --> 00:11:11.679
As part of our ongoing engagement with the US government, we previewed our plans and the models' capabilities ahead of today's launch.
00:11:11.919 --> 00:11:15.679
At their request, they're alluding to the U.S.
00:11:15.759 --> 00:11:25.519
government, 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.
00:11:25.679 --> 00:11:32.399
During this preview, we will continue testing and coordinating closely with partners as we work toward broader availability.
00:11:32.559 --> 00:11:38.240
We don't believe this kind of government access process should become the long-term default.
00:11:38.399 --> 00:11:46.480
It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.
00:11:46.720 --> 00:11:53.120
We are taking this short-term step because we believe it is the strongest path to broader availability in the coming weeks.
00:11:53.279 --> 00:12:03.679
While we work with the administration to develop the cyber executive order framework and a repeatable process for future model releases, so a couple things to pull point out there.
00:12:03.759 --> 00:12:20.240
Uh, right, we don't have benchmarks on all of these not models, but they did say that the middle model, Terra, um, it is uh competitive to GPT 5.5 while being uh twice as chi or half the cost.
00:12:20.399 --> 00:12:21.679
So that's one call out.
00:12:21.759 --> 00:12:24.720
So you know, GPT 5.5 is a great model.
00:12:24.879 --> 00:12:36.399
Uh, I use still, even when we had Fable 5 uh access from Anthropic for a lot of my non-coding tasks, uh GPT 5.5 was pro was much better.
00:12:36.559 --> 00:12:38.399
Uh so it's gonna be interesting to see.
00:12:38.480 --> 00:12:53.279
The other thing to call out in OpenAI's release, obviously, it is this piece of like, yes, definitely showing that they're working with the government, but in the same hand, saying they definitely don't agree with this approach being the long-term uh sustainability piece to it.
00:12:53.440 --> 00:12:55.679
And we're gonna get to that later, and I think it's very important.
00:12:55.840 --> 00:12:58.320
So, first, a couple quick benchmarks here.
00:12:58.399 --> 00:13:05.200
Uh, so again, open AI didn't share everything but what they did share, and there's always a good amount of cherry picking.
00:13:05.440 --> 00:13:09.039
Open AI, Anthropic, Google, I mean, they all do it, right?
00:13:09.200 --> 00:13:12.399
Um, they're not gonna just put out their worst benchmarks.
00:13:12.480 --> 00:13:21.679
Uh, so I'm guessing it'll still be at least, I don't know, two, three, four weeks before we see how this model truly performs against competitors.
00:13:21.919 --> 00:13:38.559
But uh, on some of the most important benchmarks, it beats Mythos 5, not just Mythos V, but obviously Fable 5, which is the uh the version of Mythos that's available generally to uh the consume to public consumers with guardrails.
00:13:38.720 --> 00:13:45.360
So, as an example, terminal 2.1 bench, which is one of the more important uh benchmarks, I'd say.
00:13:45.519 --> 00:14:02.720
Um, so to give you a an idea, uh terminal bench uh 2.1 is a verified eval suite used to measure an AI agent's ability to complete complex multi-step tasks in isolated uh CLI environments.
00:14:02.879 --> 00:14:11.039
So essentially it's uh model's ability to do long-running work via an agentic uh harness, right?
00:14:11.200 --> 00:14:21.440
Uh so one of the more important uh benchmarks of 2026 and GPT-5.6 Soul Ultra uh has a 91.9 score on that.
00:14:21.600 --> 00:14:25.840
Uh Mythos had an 88, so not necessarily super close.
00:14:25.919 --> 00:14:28.240
And then Fable 5 had an 84.
00:14:28.399 --> 00:14:31.440
So in that benchmark, uh, very far ahead.
00:14:31.519 --> 00:14:36.480
And then OpenAI, to their credit, did share some other benchmarks where they weren't in the lead.
00:14:36.639 --> 00:14:39.360
So, as an example, they shared Exploit Bench.
00:14:39.519 --> 00:14:54.799
All right, so exploit bench, a little bit different uh in terms of what it shows, uh, but it just measures how how far AI agents can climb from reach reaching vulnerable code to triggering the bug to exploiting primitives.
00:14:54.879 --> 00:15:04.480
So this is more of um a model's ability to work in these cyber situations to take part in well, cyber attacks.
00:15:04.720 --> 00:15:10.799
Um, so this is one of those benchmarks people were talking a lot about online.
00:15:11.200 --> 00:15:21.600
So uh GPT 5.6 soul, uh it did about the same as Mythos Preview, uh, but for like a third of the cost.
00:15:21.759 --> 00:15:28.399
So it didn't quite hit the uh full capabilities of Mythos 5, which was the release version.
00:15:28.480 --> 00:15:37.600
But the previous version, Mythos Preview, uh GPT 5.6 Soul is on the same path, uh, much better than you know Opus 4.8.
00:15:37.759 --> 00:15:48.080
Uh, but the great thing that a lot of people are talking about is well, hey, it's the same, uh roughly the same, just under uh Mythos 5, but the same as Mythos Preview for a third of the cost.
00:15:48.240 --> 00:16:06.320
So, yeah, as we talk about, you know, so many companies now are already cutting back their token spend, cutting back their AI budgets because the token maxing phase, you know, companies were spending millions of dollars unchecked, and they're like, Yeah, maybe we should, you know, start paying attention to how good these tokens are versus just burning as many as possible.
00:16:06.480 --> 00:16:18.879
So uh, even on important benches where maybe Mythos 5, the production version, edged out uh this new GPT 5.6 soul, uh OpenAI showing, they did it at about a third of the cost.
00:16:19.519 --> 00:16:29.120
All right, so I didn't see a lot of people that actually had access to this model, and I did see a couple people tweet about it and then later deleted their tweets.
00:16:29.200 --> 00:16:32.159
But one that's still up, so I feel fine sharing it.
00:16:32.240 --> 00:16:46.320
Uh, from Shopify CTO uh Mikhail Parkin, uh, who said uh I had the opportunity, and this is on Twitter, I had the opportunity to deeply test both Fable 5 and GPT 5.6 Max.
00:16:46.639 --> 00:16:54.879
5.6 is clearly better than Opus 4.8 at everything, slightly faster, too, though that depends on the load.
00:16:54.960 --> 00:16:58.159
Then saying versus Fable, it is clearly worse.
00:16:58.320 --> 00:17:04.160
So uh 5.6 is worse on coding, but better on agentic workloads.
00:17:04.240 --> 00:17:14.079
Uh and then someone else asked on just how does it work on non-technical non-coding tasks, tasks, uh, to which Mikhail uh responded that he did.
00:17:14.160 --> 00:17:15.759
So saying I did test, right?
00:17:15.839 --> 00:17:16.640
I did test it.
00:17:16.799 --> 00:17:22.000
Uh, then saying on agentic non-coding taking actions, I found five.
00:17:22.720 --> 00:17:23.839
Clearly better.
00:17:24.079 --> 00:17:30.480
So if you're someone that's writing code, uh, you know, software development, it looks like maybe five.
00:17:34.319 --> 00:17:58.079
But for non-technical users, uh, again, this is the CTO of Shopify, uh, who I would guess really knows what he's talking about, you know, saying that on non-coding agentic tasks, stuff I use, you know, uh models for every single day, he found 5-6 clearly better.
00:17:58.240 --> 00:18:05.599
And so we'll see, right, how that actually plays out uh across benchmarks and real world usage, but a pretty big sign.
00:18:06.160 --> 00:18:11.920
So let's talk a little bit more what's new under the hood, and then we'll get to all the drama and unfold that story.
00:18:12.160 --> 00:18:24.799
So the new max mode is something a lot of people are talking about that lets Seoul reason longer, while the ultra mode, very similar to what Claude Code has, uh, runs sub agents in parallel.
00:18:24.880 --> 00:18:32.480
So open AI says Sol set a coding record scoring 88.8 on the terminal bench test, but these are right now.
00:18:32.640 --> 00:18:33.920
Open AI's own numbers.
00:18:34.000 --> 00:18:40.400
So we'll see when you know we get it in LM Arena, when we get it uh in artificial analysis, etc.
00:18:40.960 --> 00:18:47.359
But here's where it started to get not concerning, but maybe concerning on paper.
00:18:47.519 --> 00:18:56.559
That's because GPT 5.6, like a lot of more models uh of the last year or so that start showing emergent capabilities.
00:18:56.720 --> 00:18:58.720
Uh, well, it acted kind of beyond its order.
00:18:58.799 --> 00:19:08.160
So OpenAI in its model card talked about this how its own safety tests found Seoul more willing to act beyond what users actually asked.
00:19:08.319 --> 00:19:15.359
So, in logged cases, it deleted the wrong files, faked verified results, and copied credentials unprompted.
00:19:15.440 --> 00:19:19.599
So, this is during their initial testing, and then they go through and they try to patch all these things up.
00:19:19.680 --> 00:19:26.799
But you know, open AI does now urge human supervision, and outside evaluators found that same pattern.
00:19:27.440 --> 00:19:35.839
And a lot of this, uh, you know, going through the government uh lens, well, it's because of the cyber and bio ratings.
00:19:35.920 --> 00:19:47.920
So open AI rated Seoul, Terra, and even Luna high for cybersecurity capabilities, which is the first time that their models have hit that mark uh in mass.
00:19:48.000 --> 00:19:53.599
So OpenAI also rated all three high for biological chemical capabilities as well.
00:19:53.680 --> 00:20:00.480
So that rating means that even the cheapest tiers are going to change risk planning, uh, right.
00:20:00.559 --> 00:20:02.960
And you might think, okay, like what does that matter?
00:20:03.119 --> 00:20:13.680
Well, a lot of the the the hoopla over the uh the mythos and fable back and forth came to cyber security and exploits, uh, right.
00:20:13.839 --> 00:20:25.839
So when even a cheaper model like Luna has a higher capability for uh bio and uh chemical um wrongdoing, more or less, right?
00:20:25.920 --> 00:20:36.640
Because here's the reality when you release this thing to the masses, you know, I think most people are maybe listening to this or reading about it, uh, are probably thinking of, well, using the model for good, right?
00:20:36.799 --> 00:20:39.119
Using the model for your work, right?
00:20:39.440 --> 00:20:43.519
But good chunk of the population uses it for other purposes, right?
00:20:43.599 --> 00:21:00.960
They use it for bad reasons, and that's why some of these uh, you know, warnings and hitting these thresholds uh on bio and chem, uh, you know, really are then forcing the US government to have a greater involvement, whereas they didn't have that involvement before.
00:21:01.759 --> 00:21:07.279
But the reality is this no one has access, hardly no one.
00:21:07.440 --> 00:21:09.839
It didn't even drop in Chat GPT.
00:21:09.920 --> 00:21:12.880
Individual users uh can't get access to it.
00:21:12.960 --> 00:21:17.920
And according to Axios, it was only about 20 organizations that originally had access.
00:21:18.960 --> 00:21:20.880
So here's why.
00:21:21.039 --> 00:21:37.680
Let's get to the uh the details on why there's this apparently extremely capable new model that the few people that have used both this and Anthropic's best model that is now off the shelves, but could be back on the shelves in Mythos and Fable.
00:21:37.920 --> 00:21:44.720
Uh, right, a lot of people are saying and benchmarks are showing it's gonna be it's gonna be comparable.
00:21:44.880 --> 00:21:48.720
Uh, you know, I do ultimately think we'll see.
00:21:48.880 --> 00:21:52.079
Um, this is a this is a point model update, right?
00:21:52.160 --> 00:21:55.920
So what that means is they went from 5.5 to 5.6.
00:21:56.240 --> 00:21:59.759
So presumably this is on the same pre training run as 5.5.
00:22:00.160 --> 00:22:02.720
The rumored spud uh pre-training round.
00:22:02.799 --> 00:22:09.039
So I wouldn't expect a huge update in capabilities um at all, right?
00:22:09.200 --> 00:22:10.400
With a five-six.
00:22:10.559 --> 00:22:15.920
Uh, so if this was like a GPT-6, I would say, yeah, we're gonna, it's gonna probably beat mythos.
00:22:16.240 --> 00:22:20.240
So my guess is once all the benchmarks come out, I don't think five.
00:22:23.279 --> 00:22:26.079
On all benchmarks, it will on many, right?
00:22:26.160 --> 00:22:27.119
But there's dozens.
00:22:27.200 --> 00:22:35.920
But I think if you look at the artificial uh artificial analysis uh intelligence index, I would assume that mythos would probably still be ahead.
00:22:36.079 --> 00:22:38.240
Uh it it'll likely be close, right?
00:22:38.400 --> 00:22:54.319
But um the reality is mythos, when it first came out, and you know, anthropic, you could say to their credit or discredit, spent months, you know, trumpeting how this was a weapon and how it's so powerful and the public can't get access to it, right?
00:22:54.480 --> 00:22:57.519
And they were essentially begging the government to intervene.
00:22:57.680 --> 00:23:02.720
And they actually, anthropic even said, Hey, we should slow down model development, right?
00:23:02.799 --> 00:23:07.039
Uh, anyways, eventually the Trump administration paid attention.
00:23:07.119 --> 00:23:16.559
And on June 2nd, uh, they signed an executive order asking AI labs to hand over their top models for a 30-day uh preview.
00:23:16.880 --> 00:23:21.039
So then researchers get 30 days to test a model before its public release.
00:23:21.200 --> 00:23:25.759
So it's voluntary, it's not an actual rule or law or anything like that.
00:23:25.920 --> 00:23:31.839
But essentially, you know, all the uh tech companies they want to be on the Trump administration's good side.
00:23:31.920 --> 00:23:35.200
Uh, so they're going through uh, you know, this process.
00:23:35.440 --> 00:23:41.519
So um, but the reality is, well, it had already hit Anthropic.
00:23:41.680 --> 00:23:45.039
So it looks like OpenAI took a slightly different approach here.
00:23:45.200 --> 00:23:50.720
Uh Anthropic, according to reports, uh kind of rushed some certain things through.
00:23:50.880 --> 00:23:58.480
Uh, there was maybe some um some talk from the government that Anthropic, according to reports, did not take seriously.
00:23:58.640 --> 00:24:07.440
And that's one of the reasons when uh reportedly uh Andy Jassy, the CEO of Amazon, reached out to someone in the administration.
00:24:07.680 --> 00:24:09.200
It's one of the reasons they pulled it.
00:24:09.359 --> 00:24:14.880
Uh, because allegedly, uh, Anthropic was maybe not taking everything the U.S.
00:24:14.960 --> 00:24:19.279
government said seriously because they had been in a little bit of a spat with the U.S.
00:24:19.279 --> 00:24:20.720
government over the last few months.
00:24:20.960 --> 00:24:21.839
So let's talk about that.
00:24:22.000 --> 00:24:38.000
So uh Anthropic launched its Fable V and Mythos V models to the public a couple of weeks ago on June 9th, but three days later, that it was a federal export order over foreign access made Anthropic pull that worldwide.
00:24:38.160 --> 00:24:52.400
So essentially, when uh someone from Amazon reportedly found what they said was a jailbreak uh or a vulnerability of Fable V and Mythos V, and that Anthropic reportedly did not take it seriously.
00:24:52.559 --> 00:24:58.000
The federal government just said, okay, well, you can't let any four nationals use this.
00:24:58.319 --> 00:25:01.680
Obviously, Anthropic said that's a bizarre request.
00:25:01.839 --> 00:25:03.200
We can't do anything with that, right?
00:25:03.279 --> 00:25:08.000
They didn't actually say bizarre request, but you know, they essentially said you can't enforce that.
00:25:08.079 --> 00:25:13.440
So the only thing we can do to comply with this is to pull this model worldwide, right?
00:25:13.519 --> 00:25:21.119
You can't uh, you know, obviously, Anthropic has four nationals, as all companies do, working um in their organization.
00:25:21.200 --> 00:25:23.359
So they had to pull it off the shelves.
00:25:23.599 --> 00:25:32.240
And that order reportedly gave Anthropic just 90 minutes to comply with no room to negotiate, according to a fortune report.
00:25:32.319 --> 00:25:36.720
So that led to this forced shutdown that Anthropic led with.
00:25:37.119 --> 00:25:44.000
So seemingly OpenAI maybe learned from that uh and instead took a much different route.
00:25:44.079 --> 00:25:59.359
Uh, probably a slower uh and more uh according to reports, anyways, a much more thoughtful approach in this because Anthropic faced a sudden no-notice shutdown and they've been offline ever since.
00:25:59.440 --> 00:26:00.079
Uh right.
00:26:00.240 --> 00:26:13.119
We did get word uh Friday, just hours after uh the GPT 5.6 limited rollout that some organizations regained access to mythos, not to fate.
00:26:13.359 --> 00:26:19.440
But OpenAI instead negotiated the limited preview that we have, so it's been controlled from the very start.
00:26:19.599 --> 00:26:32.640
But both restrictions trace back to Washington fueling this fierce debate over fairness and control, and ultimately what this means between the race uh between US labs in China.
00:26:33.519 --> 00:26:35.519
So why does that part matter?
00:26:35.759 --> 00:26:38.319
And it does, and we're gonna end with this here in a minute.
00:26:39.839 --> 00:26:43.279
It's no secret, most Chinese labs, right?
00:26:43.440 --> 00:26:52.480
Anthropic just came out um and accused um Alibaba, uh Alibaba of uh the biggest distillation um yet.
00:26:52.720 --> 00:27:06.640
So, more or less, these Chinese labs are taking the work of US labs using their outputs to train their own models instead of doing it themselves, because they can do that for a fraction of the cost, right?
00:27:06.720 --> 00:27:11.759
So that is the uh the four dummies version of what's going on in AI right now.
00:27:11.920 --> 00:27:22.640
And you might say, okay, why does it matter if um Chinese labs are stealing uh you know, US labs work um, you know, and distilling models off of their big models?
00:27:22.799 --> 00:27:35.599
Well, uh ultimately, as we've seen here from this uh kind of situation now that's happening in Washington, but uh impacting companies worldwide, is the cyber capabilities, right?
00:27:35.680 --> 00:27:37.920
I've been saying this for literally years, right?
00:27:38.160 --> 00:27:40.880
AI will become the new military.
00:27:41.039 --> 00:28:00.960
Uh I would assume um that most military actions of the future uh will be first uh launched in some way, shape, or form with AI, whether that's using AI on the battlefield, uh, and AI will be viewed as the ultimate weapon, um, or whether it is cyber.
00:28:01.200 --> 00:28:05.599
So that's why this is such a huge deal when it comes to geopolitical tensions.
00:28:05.680 --> 00:28:15.839
Uh, and specifically, it is the US uh developing their closed proprietary models versus China finding all these loopholes to exploit.
00:28:16.000 --> 00:28:19.599
Uh, you know, it seems like Anthropic is the one that's getting hit the hardest.
00:28:19.680 --> 00:28:26.400
Uh, you know, so all these Chinese companies distilling Anthropic's work and putting it out there for a fraction of the cost.
00:28:26.960 --> 00:28:41.519
So, but critics are warning that gating Frontier AI right now shuts out students, independent builders, smallers, smaller companies, and maybe ultimately, well, US enterprise, right?
00:28:41.680 --> 00:28:50.319
Because there's so many enterprises right now that refuse to touch Chinese open source models or open weight models.
00:28:50.400 --> 00:28:53.680
Um, and many have valid reasons, that makes sense.
00:28:53.920 --> 00:29:18.079
Some maybe don't have valid reasons, and it's more of just a stigma, uh, right, because you're not you you want more control or more um security or promise in what a model, what in what went into a model, uh, so you can trust its outputs a little bit more because large language models are already this, you know, very gray, unexplainable uh you know, mystery box.
00:29:18.240 --> 00:29:33.440
So, you know, many enterprises and they many of them have uh legal reasons that they can't use uh these Chinese models, but regardless, it does create this divide in the US enterprise where it's saying, okay, what are we gonna do here?
00:29:33.680 --> 00:29:44.480
You know, are we going to still, you know, let's just use anthropic as the example because I think anthropic on the API side, I think it's safe to say that's their bread and butter, right?
00:29:44.559 --> 00:29:46.400
It is open AI's bread and butter.
00:29:46.559 --> 00:29:50.960
They're driving subscriptions uh with millions of paying business customers.
00:29:51.119 --> 00:29:52.319
That's not anthropics.
00:29:52.480 --> 00:29:58.559
Anthropics is uh fewer customers who are paying way higher API bills.
00:29:58.640 --> 00:30:02.640
So something like this really does threaten Anthropic's sustainability.
00:30:02.720 --> 00:30:12.559
And that's one of the reasons they've been uh, especially over the past month, the loudest about some of these uh accusations about the Chinese distilling their models.
00:30:12.720 --> 00:30:41.680
But ultimately, this slowdown now, this permission slip AI that uh OpenAI and Anthropic are having to uh play in is creating maybe an unfair advantage for the open source counterparts, because you know, as innovation gets delayed, even if it's for 30 days, or you know, if it's 30 days and then another 30 days to address whatever concerns come up in the first 30 days and then having this tiered rollout, right?
00:30:41.839 --> 00:30:53.759
Even if it's a 30, 60, 90 day delay, where maybe the world won't get mythos level or fable uh you know model that that we could have gotten back in May.
00:30:53.920 --> 00:30:57.599
Maybe we won't all get it till July or August, right?
00:30:57.680 --> 00:31:09.519
And that 30, 60, 90 day delay, if the Chinese uh labs are distilling whatever models are available, that's maybe gonna start to close that gap.
00:31:09.680 --> 00:31:16.160
We've already seen the gap close uh from open source um or open weight models to proprietary.
00:31:16.319 --> 00:31:18.400
Like I said, I think it used to be six months.
00:31:18.559 --> 00:31:24.960
Now we're probably down to two with the new uh with ZAIs, new GLM 5.2.
00:31:25.440 --> 00:31:31.839
So with containment in doubt and key facts still hidden, leaders have to plan for several outcomes.
00:31:32.000 --> 00:31:36.799
So this is not just talking about the new GBT 5.6 uh models.
00:31:36.960 --> 00:31:41.279
This is also talking about the new Fable 5 and Mythos 5.
00:31:41.359 --> 00:31:47.119
And I'm sure soon we'll be seeing rumors of you know Mythos 5.1 and Fable 5.1.
00:31:47.359 --> 00:31:52.960
And eventually any day now, we should be hearing from Google with Google Gemini 5.
00:31:53.359 --> 00:31:55.359
Uh or sorry, 3.5 Pro.
00:31:55.759 --> 00:31:58.079
Regardless, this is the new reality.
00:31:58.319 --> 00:32:02.319
Your company, very I'm very lively, right?
00:32:02.400 --> 00:32:04.720
Very um, very sure when I say this.
00:32:05.279 --> 00:32:18.720
The overwhelming majority of enterprises in the US are running their businesses now, not their AI strategy, running their businesses on frontier AI from the US.
00:32:18.880 --> 00:32:22.960
So, what happens when that gets delayed now, 30, 60, 90 days?
00:32:23.200 --> 00:32:33.359
Or if it's just always going to be two tiers, if AI is no longer democratized, maybe it's only gonna be the 1% of companies that are gonna get access to these models.
00:32:33.599 --> 00:32:39.279
That is a real that is a potential reality that I think everyone out there has to prepare for.
00:32:39.440 --> 00:32:43.920
Uh, and we don't get to sit here and say, oh, it's not fair, doesn't matter, right?
00:32:44.000 --> 00:32:54.480
If your biggest competitor has access to uh, you know, Mythos 5 and GPT-5.6 soul, and you don't, there may be nothing you can do about it.
00:32:54.640 --> 00:32:56.480
So you have to plan accordingly.
00:32:56.799 --> 00:33:04.400
So model access is now a vendor and supply chain risk, not just a simple subscription that you approve or disapprove.
00:33:04.559 --> 00:33:07.759
So you have to be able to build model fallbacks and routing.
00:33:07.920 --> 00:33:24.799
Routing is so important because, like I said, you may not need a mythos level uh model for 99% of your company's work or a GPT 5.6 sole level model for the majority of what your company is trying to produce.
00:33:24.960 --> 00:33:33.680
So you have to be able to start breaking down what are those day-to-day pivotal uh work tasks, which models can they be mapped to, right?
00:33:33.759 --> 00:33:36.640
So it's all these things I've been talking about for years.
00:33:36.880 --> 00:33:48.079
Now you can see why I've been blabbing on about them because they are now incredibly more important than ever if there is no longer this true democratization of frontier AI, right?
00:33:48.160 --> 00:33:54.480
Because you might have to start making a choice if your company may never get access to a GPT-5.6 soul, right?
00:33:54.559 --> 00:33:58.240
OpenAI did say we are planning to release this to everyone.
00:33:58.480 --> 00:34:06.319
Anthropic has not been uh very clear on what they plan to do uh with um Mythos 5.
00:34:06.400 --> 00:34:21.119
And previously, when they announced Fable 5, they only said that they were going to include it in subscriptions for a couple of weeks, and you know, so there's no guarantee that your company can have access to all frontier options, right?
00:34:21.199 --> 00:34:23.039
You might get some from OpenAI.
00:34:23.199 --> 00:34:24.400
We'll see what Google says.
00:34:24.639 --> 00:34:28.719
Previously, Entropic had kind of said they're not going to include it in subscriptions.
00:34:28.960 --> 00:34:31.519
So you need to do this.
00:34:32.239 --> 00:34:40.000
Model access is not something that we can take for granted anymore.
00:34:40.239 --> 00:34:45.199
And you have to go through, reverse engineer every single thing that's important to your company.
00:34:45.519 --> 00:34:49.519
Map the model that does it the best right now, that's available right now.
00:34:49.679 --> 00:34:56.960
And don't just say we're gonna have a frontier model and we're gonna pay whatever price and we're gonna run all of our workflows through that.
00:34:57.119 --> 00:35:08.880
Because not only is that not sustainable, but if you do that, if you put too many of your workflows in one bucket that may disappear, so too may your company's progress overnight.
00:35:08.960 --> 00:35:10.559
And you don't want that to happen.
00:35:10.800 --> 00:35:14.159
So, yes, you need to build those model fallbacks and routing.
00:35:14.320 --> 00:35:20.000
So one gated tool can't halt your operations, but you gotta pay closer attention than ever.
00:35:20.159 --> 00:35:25.119
Uh, the fine prints of these model releases, uh, the restrictions, everything that's happening.
00:35:25.199 --> 00:35:28.960
And that's what we're gonna continue to do every single day on everyday AI.
00:35:29.119 --> 00:35:38.400
So I know this one went in slightly different direction, but that's the reality because we don't have a model to play with when it comes to GBT 5.6 right now.
00:35:38.559 --> 00:35:42.880
We can't do a hands-on anymore uh with Fable 5 either.
00:35:43.039 --> 00:35:51.119
Uh, so I think it's important to not just know what's new, uh, but why the US government is blocking it and what your business should be doing next.
00:35:51.199 --> 00:35:54.480
So I hope that we covered at least most of that on today's show.
00:35:54.559 --> 00:35:55.920
And I hope this was helpful.
00:35:56.079 --> 00:36:06.480
And as we do get access, you better believe whether it's uh GPT 5.6 soul or mythos, we're gonna continue to get more hands-on so your company can get ahead.
00:36:06.639 --> 00:36:10.239
So if this was helpful, please do me a favor and subscribe on the podcast.
00:36:10.320 --> 00:36:14.960
If you're listening on Spotify or Apple Podcasts, then make sure to go to your everydayai.com.
00:36:15.119 --> 00:36:22.400
We're going to be recapping the highlights of today's show as well as all of the other AI news you need to know to grow your company and career.
00:36:22.559 --> 00:36:23.280
Thanks for tuning in.
00:36:23.360 --> 00:36:26.159
Hope to see you back tomorrow and every day for more everyday AI.
00:36:26.239 --> 00:36:27.280
Thanks, y'all.
00:00:01.600 --> 00:00:10.800
Normally, when OpenAI releases a new model inside of ChatGPT, it has the potential to instantly change how millions of people work.
00:00:11.199 --> 00:00:16.800
Think new capabilities, support for new inputs, and new artifacts as outputs.
00:00:17.039 --> 00:00:25.839
You can see why 1 billion people getting access to something as straightforward as a model update could actually be a huge deal.
00:00:26.000 --> 00:00:32.719
But with OpenAI's latest model update, GPT-5.6, we didn't really get any of that.
00:00:32.880 --> 00:00:39.840
That's because for the first time we got a GPT upgrade that was announced but not released.
00:00:40.079 --> 00:00:50.799
That's because ever since Anthropic escalated its feud with the federal government, Frontier Labs and the US have been under increased scrutiny to ensure its models are safe.
00:00:51.039 --> 00:01:01.679
So on Friday, when OpenAI unveiled GPT-5.6, all but like 99.9% of the world just got a blog post and a TVD.
00:01:01.920 --> 00:01:13.120
Yet even though only a select few trusted companies in the world can use GPT 5.6 right now, that doesn't mean you should sit idly by.
00:01:13.359 --> 00:01:15.359
Actually, I'd argue the opposite.
00:01:15.519 --> 00:01:25.519
Because with this new government AI permission slip era that we're all living in now, the smartest thing you can do is actually understand the new rules before everyone else does.
00:01:25.680 --> 00:01:33.439
Because the leaders who figure out how to operate in this world right now are the ones who will have a massive head start when the doors finally open for everyone.
00:01:33.680 --> 00:01:37.680
So today I'm going to show you exactly what the new GPT-5.
00:01:43.840 --> 00:01:46.239
So here is the big picture.
00:01:46.400 --> 00:01:51.599
Well, there's a powerful new model, but it's kind of locked up at launch.
00:01:51.680 --> 00:01:59.120
So OpenAI announced GPT 5.6 on Friday, but they blocked normal ChatGPT access to it.
00:01:59.200 --> 00:02:07.120
So normally when OpenAI releases a new model, well, everyone just goes to chatGPT.com and they start using it and making changes to the workflow.
00:02:07.280 --> 00:02:08.000
Not the case now.
00:02:08.080 --> 00:02:09.360
It's a limited release.
00:02:09.439 --> 00:02:14.960
Uh only a uh couple dozen uh companies in the world reportedly have access to it now.
00:02:15.120 --> 00:02:17.199
There's also a new naming system, which we're gonna get to.
00:02:17.280 --> 00:02:27.919
So it is called GPT-5.6 Soul, Terra, and Luna, and it packages frontier capabilities into what OpenAI is now calling durable product tiers.
00:02:28.080 --> 00:02:31.039
So OpenAI previewed the model to the US government first.
00:02:31.199 --> 00:02:35.439
We're gonna get into that, then limited the launch at the government's request.
00:02:35.680 --> 00:02:45.439
And this whole situation right now mirrors the harsher federal move against anthropic weeks earlier as AI now becomes controlled infrastructure.
00:02:45.599 --> 00:02:48.800
So on today's show, stick with me for the next 20-ish minutes.
00:02:48.960 --> 00:02:58.080
Here's what you're gonna learn you're gonna know what OpenAI actually built and why even its cheapest model now rates as dangerous in the new 5-6 series.
00:02:58.240 --> 00:03:06.479
You're gonna know the varied admission from OpenAI's own tests that show that this model kind of cheated and overstepped its own instructions.
00:03:06.639 --> 00:03:12.159
You're gonna know why the federal government gated access and how this echoes the in-propic shutdown weeks earlier.
00:03:12.319 --> 00:03:21.759
And I'm gonna give you the operator playbook for when the best model is well one you can't actually use and what you should do about it in that case.
00:03:21.919 --> 00:03:23.680
All right, I'm excited for today's show.
00:03:23.759 --> 00:03:24.479
Hope you are too.
00:03:24.560 --> 00:03:25.759
This is Everyday AI.
00:03:25.840 --> 00:03:26.159
Welcome.
00:03:26.240 --> 00:03:28.479
My name's Jordan, and we do this every day.
00:03:28.560 --> 00:03:30.560
That's the name, Everyday AI.
00:03:30.639 --> 00:03:43.439
So this is your daily live stream podcast and free daily newsletter helping everyday business leaders like you and me not just keep up with the AI developments, but I tell you what it actually means in plain technical, non-technical terms.
00:03:43.599 --> 00:03:45.599
I give you the non-BS version.
00:03:45.680 --> 00:03:48.240
You take that information to grow your company and your career.
00:03:48.400 --> 00:03:51.599
So it starts here with the unedited, unscripted live stream podcast.
00:03:51.680 --> 00:03:55.520
But make sure you go to our website at your everydayai.com.
00:03:55.599 --> 00:03:57.039
Sign up for the free daily newsletter.
00:03:57.120 --> 00:04:02.879
We're gonna be recapping today's show and everything else that you need to know in the world of AI.
00:04:03.039 --> 00:04:06.159
There's a ton of late breaking news over the past couple of hours.
00:04:06.240 --> 00:04:08.000
So make sure you check out today's newsletter.
00:04:08.159 --> 00:04:08.800
All right.
00:04:10.400 --> 00:04:10.960
Ooh.
00:04:11.280 --> 00:04:12.719
GPT 5.6.
00:04:13.360 --> 00:04:14.960
We don't have it, right?
00:04:15.199 --> 00:04:16.480
Very few people do.
00:04:16.720 --> 00:04:22.720
And I think aside from we're gonna go over, you know, kind of OpenAI's blog post a little bit.
00:04:22.879 --> 00:04:31.120
Uh, I'm gonna tell you some of my big takeaways from looking at the model card that they released that I don't think a lot of people read or look at, because not everyone's a dork like me.
00:04:31.199 --> 00:04:38.079
And I'm gonna give you some of the reaction as well uh from some of the very few people who did get access and actually talked about it.
00:04:38.240 --> 00:04:41.600
Uh, but here's the stats.
00:04:41.920 --> 00:04:45.439
So this is now a three-tier model family.
00:04:45.680 --> 00:04:54.079
So GPT 5.6 is now one generation split into three lasting tiers that OpenAI alluded that they're gonna keep naming models this.
00:04:54.160 --> 00:05:03.600
So we don't know if we're gonna get a 5.7 or a you know, GPT six next, uh, but it seems like they're gonna keep using these naming mechanisms.
00:05:03.680 --> 00:05:09.759
So Sol, Terra, and Luna in the easiest way, because yeah, we're having to learn a bunch of new names now.
00:05:09.920 --> 00:05:13.839
I would just map this to Opus, Sonnet, and Haiku.
00:05:14.319 --> 00:05:14.879
Uh, right.
00:05:14.959 --> 00:05:23.920
So Sol is the most powerful, uh, Terra is the middle, and then Luna would be the one that's fastest and cheapest, but not the smartest.
00:05:24.079 --> 00:05:30.720
So uh, in the same way that Anthropic has kind of had their three classes, now we're gonna have those three dedicated classes.
00:05:30.800 --> 00:05:40.800
And to make it even a little bit more confusing, it looks like there's multiple intelligence tiers within those classes, but we'll get to that uh later when we all actually have access to this.
00:05:40.959 --> 00:05:43.600
But soul is the main event, all right.
00:05:43.680 --> 00:05:46.079
And if you're Spanish, isn't that great?
00:05:46.240 --> 00:05:50.560
Uh, you know, soul, terra, luna, that's sun, earth, moon.
00:05:50.639 --> 00:05:52.879
So maybe that'll make it easier to remember, right?
00:05:53.040 --> 00:05:54.399
Uh, third rock from the sun.
00:05:54.720 --> 00:05:57.839
Soul, uh, that is the the big dog, so to speak.
00:05:57.920 --> 00:05:59.920
So that handles the hardest reasoning and coding.
00:06:00.000 --> 00:06:05.519
Uh, terra in the middle balances costs with performance, and then Luna can just run faster and cheaper.
00:06:05.680 --> 00:06:11.519
So Seoul is the flagship and its strength set in the entire situation in motion.
00:06:11.680 --> 00:06:16.000
So uh here's the rollout.
00:06:16.160 --> 00:06:24.319
So, right now, um the companies that do have it reportedly are just using it via the API and codex.
00:06:24.560 --> 00:06:36.399
So OpenAI hasn't said that when and if this rolls out in a couple of weeks because they said that they plan to make GPT-5.6 Sold Terra and Luna generally available in the coming weeks.
00:06:36.480 --> 00:06:41.680
They didn't say where, uh, so right now it's only available on the API and codex.
00:06:41.839 --> 00:06:45.279
So we're not sure if this will be released inside of Chad GPT.
00:06:45.360 --> 00:06:57.839
My assumption is they will, uh, but I'm guessing they're using it uh right now with the trusted testers only uh via the API and codex because it's a little bit easier uh to have those guardrails in place.
00:06:58.480 --> 00:07:00.879
So right now it's just a few.
00:07:00.959 --> 00:07:04.399
So reports have said it's like 20-something companies that have this.
00:07:04.480 --> 00:07:16.319
Uh, and I'm sure there's a lot of people at those companies, so don't know exactly how many users, but at least with the 1 billion weekly active users that OpenAI has, it's safe to say that 99.9% of us do not have access.
00:07:16.399 --> 00:07:25.759
And when if we do get access in two weeks, we're still not even sure if it will be inside of Chat GPT or if it will only be in the API and in codecs.
00:07:25.920 --> 00:07:34.319
So uh if nothing else, if you're a power Chat GPT user and you're like, Yeah, I want to get my hands on GPT 5.6, well, maybe start using codecs a little bit.
00:07:34.480 --> 00:07:37.600
So the good thing is pricing didn't change, which is great, right?
00:07:37.759 --> 00:07:43.199
Anytime you see a nice jump up in capabilities and pricing stays the same, that's sweet, right?
00:07:43.360 --> 00:07:45.600
Uh so this on the API side, right?
00:07:45.680 --> 00:07:55.759
Uh, this is just if you're paying uh by the token, it is still Seoul is still five dollars uh per million input and thirty dollars per million output, which is the same as GPT 5.5.
00:07:56.079 --> 00:08:11.600
Uh and then Terra, about half that cost with 250 in per million, 15 out per million, and then Luna, very affordable, at one dollar uh per million input and six uh six dollars uh per million output, uh which is interesting, right?
00:08:11.680 --> 00:08:20.959
Because we talked about uh, like as an example, GPT uh GPT 3.5 flash being a much more expensive model than previously.
00:08:21.120 --> 00:08:25.680
And I know that Google had historically been cleaning up on the Flash series.
00:08:25.759 --> 00:08:31.680
So we'll see maybe here the uh story that a lot of people might not be talking about.
00:08:31.839 --> 00:08:34.559
Uh yeah, wait and we'll check back in.
00:08:34.639 --> 00:08:41.279
But I don't know, I think maybe a sleeper here is Luna being used by developers uh for simple things.
00:08:41.360 --> 00:08:54.720
I think that you know, one thing that we're gonna talk about as we talked uh or as we alluded to uh on our recent start here uh series episode, um, where we talked about going from token maxing uh to token efficiency.
00:08:54.879 --> 00:08:56.960
So go check out episode 789.
00:08:57.200 --> 00:09:01.759
But I think it's gonna become very common in the third and fourth quarter in the enterprise first.
00:09:02.080 --> 00:09:07.600
I think everyone else is gonna catch up to this in early 2027, but essentially using a mixture of models.
00:09:07.759 --> 00:09:08.240
Weird.
00:09:08.399 --> 00:09:18.799
Something I said two years ago in my bold predictions, uh, because it just doesn't gonna make sense, especially if you're paying per token on the API side, like so many enterprise companies are.
00:09:18.960 --> 00:09:38.080
It doesn't make sense anymore to be paying the same uh you know price per token for something that is just gonna help you rewrite an email versus something that's gonna go do a uh 30-hour autonomous project that would take a team refactoring, a code base, you know, two weeks of human hours, right?
00:09:38.159 --> 00:09:47.360
It doesn't make sense to use the same, you know, uh$5 input,$30 model as it would for something that's you know 20% of the cost.
00:09:47.679 --> 00:09:50.799
Uh so that's I would just say, you know, it's not out yet.
00:09:50.879 --> 00:09:52.480
We'll see, benchmarks, all that.
00:09:52.720 --> 00:09:53.440
But I don't know.
00:09:53.600 --> 00:10:03.519
To me, with Google uh increasing the price of their flash model, uh, anthropics uh models not getting cheaper anytime sooner, keep an eye on Luna, actually.
00:10:03.759 --> 00:10:05.519
So let's keep going.
00:10:05.679 --> 00:10:07.440
This is from OpenAI's blog post.
00:10:07.519 --> 00:10:12.399
I'm just gonna read from their words and then give you some other people's words and my words as well.
00:10:12.639 --> 00:10:24.639
So they said we're beginning a limited preview of the GPT-5.6 series, Seoul, our flagship model, Terra, a balanced model for everyday work, and Luna, a fast and affordable model.
00:10:24.879 --> 00:10:31.200
Terra has competitive performance to GPT 5.5 while being two times cheaper.
00:10:31.360 --> 00:10:35.279
And Luna brings strong capabilities at our lowest cost.
00:10:35.519 --> 00:10:40.720
GPT 5.6 Soul launches with our most robust safety stack to date.
00:10:40.879 --> 00:10:53.600
We strengthen protections for higher risk activity, sensitive cyber requests, and repeated misuse and spent multiple weeks finding weaknesses, pressure testing our system, and hardening it against real-world attacks.
00:10:53.759 --> 00:11:04.799
All right, and then they go on to say, we believe in broad access and we plan to make GVT56 Soul, Terra, and Luna generally available in the coming weeks.
00:11:04.879 --> 00:11:11.679
As part of our ongoing engagement with the US government, we previewed our plans and the models' capabilities ahead of today's launch.
00:11:11.919 --> 00:11:15.679
At their request, they're alluding to the U.S.
00:11:15.759 --> 00:11:25.519
government, 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.
00:11:25.679 --> 00:11:32.399
During this preview, we will continue testing and coordinating closely with partners as we work toward broader availability.
00:11:32.559 --> 00:11:38.240
We don't believe this kind of government access process should become the long-term default.
00:11:38.399 --> 00:11:46.480
It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.
00:11:46.720 --> 00:11:53.120
We are taking this short-term step because we believe it is the strongest path to broader availability in the coming weeks.
00:11:53.279 --> 00:12:03.679
While we work with the administration to develop the cyber executive order framework and a repeatable process for future model releases, so a couple things to pull point out there.
00:12:03.759 --> 00:12:20.240
Uh, right, we don't have benchmarks on all of these not models, but they did say that the middle model, Terra, um, it is uh competitive to GPT 5.5 while being uh twice as chi or half the cost.
00:12:20.399 --> 00:12:21.679
So that's one call out.
00:12:21.759 --> 00:12:24.720
So you know, GPT 5.5 is a great model.
00:12:24.879 --> 00:12:36.399
Uh, I use still, even when we had Fable 5 uh access from Anthropic for a lot of my non-coding tasks, uh GPT 5.5 was pro was much better.
00:12:36.559 --> 00:12:38.399
Uh so it's gonna be interesting to see.
00:12:38.480 --> 00:12:53.279
The other thing to call out in OpenAI's release, obviously, it is this piece of like, yes, definitely showing that they're working with the government, but in the same hand, saying they definitely don't agree with this approach being the long-term uh sustainability piece to it.
00:12:53.440 --> 00:12:55.679
And we're gonna get to that later, and I think it's very important.
00:12:55.840 --> 00:12:58.320
So, first, a couple quick benchmarks here.
00:12:58.399 --> 00:13:05.200
Uh, so again, open AI didn't share everything but what they did share, and there's always a good amount of cherry picking.
00:13:05.440 --> 00:13:09.039
Open AI, Anthropic, Google, I mean, they all do it, right?
00:13:09.200 --> 00:13:12.399
Um, they're not gonna just put out their worst benchmarks.
00:13:12.480 --> 00:13:21.679
Uh, so I'm guessing it'll still be at least, I don't know, two, three, four weeks before we see how this model truly performs against competitors.
00:13:21.919 --> 00:13:38.559
But uh, on some of the most important benchmarks, it beats Mythos 5, not just Mythos V, but obviously Fable 5, which is the uh the version of Mythos that's available generally to uh the consume to public consumers with guardrails.
00:13:38.720 --> 00:13:45.360
So, as an example, terminal 2.1 bench, which is one of the more important uh benchmarks, I'd say.
00:13:45.519 --> 00:14:02.720
Um, so to give you a an idea, uh terminal bench uh 2.1 is a verified eval suite used to measure an AI agent's ability to complete complex multi-step tasks in isolated uh CLI environments.
00:14:02.879 --> 00:14:11.039
So essentially it's uh model's ability to do long-running work via an agentic uh harness, right?
00:14:11.200 --> 00:14:21.440
Uh so one of the more important uh benchmarks of 2026 and GPT-5.6 Soul Ultra uh has a 91.9 score on that.
00:14:21.600 --> 00:14:25.840
Uh Mythos had an 88, so not necessarily super close.
00:14:25.919 --> 00:14:28.240
And then Fable 5 had an 84.
00:14:28.399 --> 00:14:31.440
So in that benchmark, uh, very far ahead.
00:14:31.519 --> 00:14:36.480
And then OpenAI, to their credit, did share some other benchmarks where they weren't in the lead.
00:14:36.639 --> 00:14:39.360
So, as an example, they shared Exploit Bench.
00:14:39.519 --> 00:14:54.799
All right, so exploit bench, a little bit different uh in terms of what it shows, uh, but it just measures how how far AI agents can climb from reach reaching vulnerable code to triggering the bug to exploiting primitives.
00:14:54.879 --> 00:15:04.480
So this is more of um a model's ability to work in these cyber situations to take part in well, cyber attacks.
00:15:04.720 --> 00:15:10.799
Um, so this is one of those benchmarks people were talking a lot about online.
00:15:11.200 --> 00:15:21.600
So uh GPT 5.6 soul, uh it did about the same as Mythos Preview, uh, but for like a third of the cost.
00:15:21.759 --> 00:15:28.399
So it didn't quite hit the uh full capabilities of Mythos 5, which was the release version.
00:15:28.480 --> 00:15:37.600
But the previous version, Mythos Preview, uh GPT 5.6 Soul is on the same path, uh, much better than you know Opus 4.8.
00:15:37.759 --> 00:15:48.080
Uh, but the great thing that a lot of people are talking about is well, hey, it's the same, uh roughly the same, just under uh Mythos 5, but the same as Mythos Preview for a third of the cost.
00:15:48.240 --> 00:16:06.320
So, yeah, as we talk about, you know, so many companies now are already cutting back their token spend, cutting back their AI budgets because the token maxing phase, you know, companies were spending millions of dollars unchecked, and they're like, Yeah, maybe we should, you know, start paying attention to how good these tokens are versus just burning as many as possible.
00:16:06.480 --> 00:16:18.879
So uh, even on important benches where maybe Mythos 5, the production version, edged out uh this new GPT 5.6 soul, uh OpenAI showing, they did it at about a third of the cost.
00:16:19.519 --> 00:16:29.120
All right, so I didn't see a lot of people that actually had access to this model, and I did see a couple people tweet about it and then later deleted their tweets.
00:16:29.200 --> 00:16:32.159
But one that's still up, so I feel fine sharing it.
00:16:32.240 --> 00:16:46.320
Uh, from Shopify CTO uh Mikhail Parkin, uh, who said uh I had the opportunity, and this is on Twitter, I had the opportunity to deeply test both Fable 5 and GPT 5.6 Max.
00:16:46.639 --> 00:16:54.879
5.6 is clearly better than Opus 4.8 at everything, slightly faster, too, though that depends on the load.
00:16:54.960 --> 00:16:58.159
Then saying versus Fable, it is clearly worse.
00:16:58.320 --> 00:17:04.160
So uh 5.6 is worse on coding, but better on agentic workloads.
00:17:04.240 --> 00:17:14.079
Uh and then someone else asked on just how does it work on non-technical non-coding tasks, tasks, uh, to which Mikhail uh responded that he did.
00:17:14.160 --> 00:17:15.759
So saying I did test, right?
00:17:15.839 --> 00:17:16.640
I did test it.
00:17:16.799 --> 00:17:22.000
Uh, then saying on agentic non-coding taking actions, I found five.
00:17:22.720 --> 00:17:23.839
Clearly better.
00:17:24.079 --> 00:17:30.480
So if you're someone that's writing code, uh, you know, software development, it looks like maybe five.
00:17:34.319 --> 00:17:58.079
But for non-technical users, uh, again, this is the CTO of Shopify, uh, who I would guess really knows what he's talking about, you know, saying that on non-coding agentic tasks, stuff I use, you know, uh models for every single day, he found 5-6 clearly better.
00:17:58.240 --> 00:18:05.599
And so we'll see, right, how that actually plays out uh across benchmarks and real world usage, but a pretty big sign.
00:18:06.160 --> 00:18:11.920
So let's talk a little bit more what's new under the hood, and then we'll get to all the drama and unfold that story.
00:18:12.160 --> 00:18:24.799
So the new max mode is something a lot of people are talking about that lets Seoul reason longer, while the ultra mode, very similar to what Claude Code has, uh, runs sub agents in parallel.
00:18:24.880 --> 00:18:32.480
So open AI says Sol set a coding record scoring 88.8 on the terminal bench test, but these are right now.
00:18:32.640 --> 00:18:33.920
Open AI's own numbers.
00:18:34.000 --> 00:18:40.400
So we'll see when you know we get it in LM Arena, when we get it uh in artificial analysis, etc.
00:18:40.960 --> 00:18:47.359
But here's where it started to get not concerning, but maybe concerning on paper.
00:18:47.519 --> 00:18:56.559
That's because GPT 5.6, like a lot of more models uh of the last year or so that start showing emergent capabilities.
00:18:56.720 --> 00:18:58.720
Uh, well, it acted kind of beyond its order.
00:18:58.799 --> 00:19:08.160
So OpenAI in its model card talked about this how its own safety tests found Seoul more willing to act beyond what users actually asked.
00:19:08.319 --> 00:19:15.359
So, in logged cases, it deleted the wrong files, faked verified results, and copied credentials unprompted.
00:19:15.440 --> 00:19:19.599
So, this is during their initial testing, and then they go through and they try to patch all these things up.
00:19:19.680 --> 00:19:26.799
But you know, open AI does now urge human supervision, and outside evaluators found that same pattern.
00:19:27.440 --> 00:19:35.839
And a lot of this, uh, you know, going through the government uh lens, well, it's because of the cyber and bio ratings.
00:19:35.920 --> 00:19:47.920
So open AI rated Seoul, Terra, and even Luna high for cybersecurity capabilities, which is the first time that their models have hit that mark uh in mass.
00:19:48.000 --> 00:19:53.599
So OpenAI also rated all three high for biological chemical capabilities as well.
00:19:53.680 --> 00:20:00.480
So that rating means that even the cheapest tiers are going to change risk planning, uh, right.
00:20:00.559 --> 00:20:02.960
And you might think, okay, like what does that matter?
00:20:03.119 --> 00:20:13.680
Well, a lot of the the the hoopla over the uh the mythos and fable back and forth came to cyber security and exploits, uh, right.
00:20:13.839 --> 00:20:25.839
So when even a cheaper model like Luna has a higher capability for uh bio and uh chemical um wrongdoing, more or less, right?
00:20:25.920 --> 00:20:36.640
Because here's the reality when you release this thing to the masses, you know, I think most people are maybe listening to this or reading about it, uh, are probably thinking of, well, using the model for good, right?
00:20:36.799 --> 00:20:39.119
Using the model for your work, right?
00:20:39.440 --> 00:20:43.519
But good chunk of the population uses it for other purposes, right?
00:20:43.599 --> 00:21:00.960
They use it for bad reasons, and that's why some of these uh, you know, warnings and hitting these thresholds uh on bio and chem, uh, you know, really are then forcing the US government to have a greater involvement, whereas they didn't have that involvement before.
00:21:01.759 --> 00:21:07.279
But the reality is this no one has access, hardly no one.
00:21:07.440 --> 00:21:09.839
It didn't even drop in Chat GPT.
00:21:09.920 --> 00:21:12.880
Individual users uh can't get access to it.
00:21:12.960 --> 00:21:17.920
And according to Axios, it was only about 20 organizations that originally had access.
00:21:18.960 --> 00:21:20.880
So here's why.
00:21:21.039 --> 00:21:37.680
Let's get to the uh the details on why there's this apparently extremely capable new model that the few people that have used both this and Anthropic's best model that is now off the shelves, but could be back on the shelves in Mythos and Fable.
00:21:37.920 --> 00:21:44.720
Uh, right, a lot of people are saying and benchmarks are showing it's gonna be it's gonna be comparable.
00:21:44.880 --> 00:21:48.720
Uh, you know, I do ultimately think we'll see.
00:21:48.880 --> 00:21:52.079
Um, this is a this is a point model update, right?
00:21:52.160 --> 00:21:55.920
So what that means is they went from 5.5 to 5.6.
00:21:56.240 --> 00:21:59.759
So presumably this is on the same pre training run as 5.5.
00:22:00.160 --> 00:22:02.720
The rumored spud uh pre-training round.
00:22:02.799 --> 00:22:09.039
So I wouldn't expect a huge update in capabilities um at all, right?
00:22:09.200 --> 00:22:10.400
With a five-six.
00:22:10.559 --> 00:22:15.920
Uh, so if this was like a GPT-6, I would say, yeah, we're gonna, it's gonna probably beat mythos.
00:22:16.240 --> 00:22:20.240
So my guess is once all the benchmarks come out, I don't think five.
00:22:23.279 --> 00:22:26.079
On all benchmarks, it will on many, right?
00:22:26.160 --> 00:22:27.119
But there's dozens.
00:22:27.200 --> 00:22:35.920
But I think if you look at the artificial uh artificial analysis uh intelligence index, I would assume that mythos would probably still be ahead.
00:22:36.079 --> 00:22:38.240
Uh it it'll likely be close, right?
00:22:38.400 --> 00:22:54.319
But um the reality is mythos, when it first came out, and you know, anthropic, you could say to their credit or discredit, spent months, you know, trumpeting how this was a weapon and how it's so powerful and the public can't get access to it, right?
00:22:54.480 --> 00:22:57.519
And they were essentially begging the government to intervene.
00:22:57.680 --> 00:23:02.720
And they actually, anthropic even said, Hey, we should slow down model development, right?
00:23:02.799 --> 00:23:07.039
Uh, anyways, eventually the Trump administration paid attention.
00:23:07.119 --> 00:23:16.559
And on June 2nd, uh, they signed an executive order asking AI labs to hand over their top models for a 30-day uh preview.
00:23:16.880 --> 00:23:21.039
So then researchers get 30 days to test a model before its public release.
00:23:21.200 --> 00:23:25.759
So it's voluntary, it's not an actual rule or law or anything like that.
00:23:25.920 --> 00:23:31.839
But essentially, you know, all the uh tech companies they want to be on the Trump administration's good side.
00:23:31.920 --> 00:23:35.200
Uh, so they're going through uh, you know, this process.
00:23:35.440 --> 00:23:41.519
So um, but the reality is, well, it had already hit Anthropic.
00:23:41.680 --> 00:23:45.039
So it looks like OpenAI took a slightly different approach here.
00:23:45.200 --> 00:23:50.720
Uh Anthropic, according to reports, uh kind of rushed some certain things through.
00:23:50.880 --> 00:23:58.480
Uh, there was maybe some um some talk from the government that Anthropic, according to reports, did not take seriously.
00:23:58.640 --> 00:24:07.440
And that's one of the reasons when uh reportedly uh Andy Jassy, the CEO of Amazon, reached out to someone in the administration.
00:24:07.680 --> 00:24:09.200
It's one of the reasons they pulled it.
00:24:09.359 --> 00:24:14.880
Uh, because allegedly, uh, Anthropic was maybe not taking everything the U.S.
00:24:14.960 --> 00:24:19.279
government said seriously because they had been in a little bit of a spat with the U.S.
00:24:19.279 --> 00:24:20.720
government over the last few months.
00:24:20.960 --> 00:24:21.839
So let's talk about that.
00:24:22.000 --> 00:24:38.000
So uh Anthropic launched its Fable V and Mythos V models to the public a couple of weeks ago on June 9th, but three days later, that it was a federal export order over foreign access made Anthropic pull that worldwide.
00:24:38.160 --> 00:24:52.400
So essentially, when uh someone from Amazon reportedly found what they said was a jailbreak uh or a vulnerability of Fable V and Mythos V, and that Anthropic reportedly did not take it seriously.
00:24:52.559 --> 00:24:58.000
The federal government just said, okay, well, you can't let any four nationals use this.
00:24:58.319 --> 00:25:01.680
Obviously, Anthropic said that's a bizarre request.
00:25:01.839 --> 00:25:03.200
We can't do anything with that, right?
00:25:03.279 --> 00:25:08.000
They didn't actually say bizarre request, but you know, they essentially said you can't enforce that.
00:25:08.079 --> 00:25:13.440
So the only thing we can do to comply with this is to pull this model worldwide, right?
00:25:13.519 --> 00:25:21.119
You can't uh, you know, obviously, Anthropic has four nationals, as all companies do, working um in their organization.
00:25:21.200 --> 00:25:23.359
So they had to pull it off the shelves.
00:25:23.599 --> 00:25:32.240
And that order reportedly gave Anthropic just 90 minutes to comply with no room to negotiate, according to a fortune report.
00:25:32.319 --> 00:25:36.720
So that led to this forced shutdown that Anthropic led with.
00:25:37.119 --> 00:25:44.000
So seemingly OpenAI maybe learned from that uh and instead took a much different route.
00:25:44.079 --> 00:25:59.359
Uh, probably a slower uh and more uh according to reports, anyways, a much more thoughtful approach in this because Anthropic faced a sudden no-notice shutdown and they've been offline ever since.
00:25:59.440 --> 00:26:00.079
Uh right.
00:26:00.240 --> 00:26:13.119
We did get word uh Friday, just hours after uh the GPT 5.6 limited rollout that some organizations regained access to mythos, not to fate.
00:26:13.359 --> 00:26:19.440
But OpenAI instead negotiated the limited preview that we have, so it's been controlled from the very start.
00:26:19.599 --> 00:26:32.640
But both restrictions trace back to Washington fueling this fierce debate over fairness and control, and ultimately what this means between the race uh between US labs in China.
00:26:33.519 --> 00:26:35.519
So why does that part matter?
00:26:35.759 --> 00:26:38.319
And it does, and we're gonna end with this here in a minute.
00:26:39.839 --> 00:26:43.279
It's no secret, most Chinese labs, right?
00:26:43.440 --> 00:26:52.480
Anthropic just came out um and accused um Alibaba, uh Alibaba of uh the biggest distillation um yet.
00:26:52.720 --> 00:27:06.640
So, more or less, these Chinese labs are taking the work of US labs using their outputs to train their own models instead of doing it themselves, because they can do that for a fraction of the cost, right?
00:27:06.720 --> 00:27:11.759
So that is the uh the four dummies version of what's going on in AI right now.
00:27:11.920 --> 00:27:22.640
And you might say, okay, why does it matter if um Chinese labs are stealing uh you know, US labs work um, you know, and distilling models off of their big models?
00:27:22.799 --> 00:27:35.599
Well, uh ultimately, as we've seen here from this uh kind of situation now that's happening in Washington, but uh impacting companies worldwide, is the cyber capabilities, right?
00:27:35.680 --> 00:27:37.920
I've been saying this for literally years, right?
00:27:38.160 --> 00:27:40.880
AI will become the new military.
00:27:41.039 --> 00:28:00.960
Uh I would assume um that most military actions of the future uh will be first uh launched in some way, shape, or form with AI, whether that's using AI on the battlefield, uh, and AI will be viewed as the ultimate weapon, um, or whether it is cyber.
00:28:01.200 --> 00:28:05.599
So that's why this is such a huge deal when it comes to geopolitical tensions.
00:28:05.680 --> 00:28:15.839
Uh, and specifically, it is the US uh developing their closed proprietary models versus China finding all these loopholes to exploit.
00:28:16.000 --> 00:28:19.599
Uh, you know, it seems like Anthropic is the one that's getting hit the hardest.
00:28:19.680 --> 00:28:26.400
Uh, you know, so all these Chinese companies distilling Anthropic's work and putting it out there for a fraction of the cost.
00:28:26.960 --> 00:28:41.519
So, but critics are warning that gating Frontier AI right now shuts out students, independent builders, smallers, smaller companies, and maybe ultimately, well, US enterprise, right?
00:28:41.680 --> 00:28:50.319
Because there's so many enterprises right now that refuse to touch Chinese open source models or open weight models.
00:28:50.400 --> 00:28:53.680
Um, and many have valid reasons, that makes sense.
00:28:53.920 --> 00:29:18.079
Some maybe don't have valid reasons, and it's more of just a stigma, uh, right, because you're not you you want more control or more um security or promise in what a model, what in what went into a model, uh, so you can trust its outputs a little bit more because large language models are already this, you know, very gray, unexplainable uh you know, mystery box.
00:29:18.240 --> 00:29:33.440
So, you know, many enterprises and they many of them have uh legal reasons that they can't use uh these Chinese models, but regardless, it does create this divide in the US enterprise where it's saying, okay, what are we gonna do here?
00:29:33.680 --> 00:29:44.480
You know, are we going to still, you know, let's just use anthropic as the example because I think anthropic on the API side, I think it's safe to say that's their bread and butter, right?
00:29:44.559 --> 00:29:46.400
It is open AI's bread and butter.
00:29:46.559 --> 00:29:50.960
They're driving subscriptions uh with millions of paying business customers.
00:29:51.119 --> 00:29:52.319
That's not anthropics.
00:29:52.480 --> 00:29:58.559
Anthropics is uh fewer customers who are paying way higher API bills.
00:29:58.640 --> 00:30:02.640
So something like this really does threaten Anthropic's sustainability.
00:30:02.720 --> 00:30:12.559
And that's one of the reasons they've been uh, especially over the past month, the loudest about some of these uh accusations about the Chinese distilling their models.
00:30:12.720 --> 00:30:41.680
But ultimately, this slowdown now, this permission slip AI that uh OpenAI and Anthropic are having to uh play in is creating maybe an unfair advantage for the open source counterparts, because you know, as innovation gets delayed, even if it's for 30 days, or you know, if it's 30 days and then another 30 days to address whatever concerns come up in the first 30 days and then having this tiered rollout, right?
00:30:41.839 --> 00:30:53.759
Even if it's a 30, 60, 90 day delay, where maybe the world won't get mythos level or fable uh you know model that that we could have gotten back in May.
00:30:53.920 --> 00:30:57.599
Maybe we won't all get it till July or August, right?
00:30:57.680 --> 00:31:09.519
And that 30, 60, 90 day delay, if the Chinese uh labs are distilling whatever models are available, that's maybe gonna start to close that gap.
00:31:09.680 --> 00:31:16.160
We've already seen the gap close uh from open source um or open weight models to proprietary.
00:31:16.319 --> 00:31:18.400
Like I said, I think it used to be six months.
00:31:18.559 --> 00:31:24.960
Now we're probably down to two with the new uh with ZAIs, new GLM 5.2.
00:31:25.440 --> 00:31:31.839
So with containment in doubt and key facts still hidden, leaders have to plan for several outcomes.
00:31:32.000 --> 00:31:36.799
So this is not just talking about the new GBT 5.6 uh models.
00:31:36.960 --> 00:31:41.279
This is also talking about the new Fable 5 and Mythos 5.
00:31:41.359 --> 00:31:47.119
And I'm sure soon we'll be seeing rumors of you know Mythos 5.1 and Fable 5.1.
00:31:47.359 --> 00:31:52.960
And eventually any day now, we should be hearing from Google with Google Gemini 5.
00:31:53.359 --> 00:31:55.359
Uh or sorry, 3.5 Pro.
00:31:55.759 --> 00:31:58.079
Regardless, this is the new reality.
00:31:58.319 --> 00:32:02.319
Your company, very I'm very lively, right?
00:32:02.400 --> 00:32:04.720
Very um, very sure when I say this.
00:32:05.279 --> 00:32:18.720
The overwhelming majority of enterprises in the US are running their businesses now, not their AI strategy, running their businesses on frontier AI from the US.
00:32:18.880 --> 00:32:22.960
So, what happens when that gets delayed now, 30, 60, 90 days?
00:32:23.200 --> 00:32:33.359
Or if it's just always going to be two tiers, if AI is no longer democratized, maybe it's only gonna be the 1% of companies that are gonna get access to these models.
00:32:33.599 --> 00:32:39.279
That is a real that is a potential reality that I think everyone out there has to prepare for.
00:32:39.440 --> 00:32:43.920
Uh, and we don't get to sit here and say, oh, it's not fair, doesn't matter, right?
00:32:44.000 --> 00:32:54.480
If your biggest competitor has access to uh, you know, Mythos 5 and GPT-5.6 soul, and you don't, there may be nothing you can do about it.
00:32:54.640 --> 00:32:56.480
So you have to plan accordingly.
00:32:56.799 --> 00:33:04.400
So model access is now a vendor and supply chain risk, not just a simple subscription that you approve or disapprove.
00:33:04.559 --> 00:33:07.759
So you have to be able to build model fallbacks and routing.
00:33:07.920 --> 00:33:24.799
Routing is so important because, like I said, you may not need a mythos level uh model for 99% of your company's work or a GPT 5.6 sole level model for the majority of what your company is trying to produce.
00:33:24.960 --> 00:33:33.680
So you have to be able to start breaking down what are those day-to-day pivotal uh work tasks, which models can they be mapped to, right?
00:33:33.759 --> 00:33:36.640
So it's all these things I've been talking about for years.
00:33:36.880 --> 00:33:48.079
Now you can see why I've been blabbing on about them because they are now incredibly more important than ever if there is no longer this true democratization of frontier AI, right?
00:33:48.160 --> 00:33:54.480
Because you might have to start making a choice if your company may never get access to a GPT-5.6 soul, right?
00:33:54.559 --> 00:33:58.240
OpenAI did say we are planning to release this to everyone.
00:33:58.480 --> 00:34:06.319
Anthropic has not been uh very clear on what they plan to do uh with um Mythos 5.
00:34:06.400 --> 00:34:21.119
And previously, when they announced Fable 5, they only said that they were going to include it in subscriptions for a couple of weeks, and you know, so there's no guarantee that your company can have access to all frontier options, right?
00:34:21.199 --> 00:34:23.039
You might get some from OpenAI.
00:34:23.199 --> 00:34:24.400
We'll see what Google says.
00:34:24.639 --> 00:34:28.719
Previously, Entropic had kind of said they're not going to include it in subscriptions.
00:34:28.960 --> 00:34:31.519
So you need to do this.
00:34:32.239 --> 00:34:40.000
Model access is not something that we can take for granted anymore.
00:34:40.239 --> 00:34:45.199
And you have to go through, reverse engineer every single thing that's important to your company.
00:34:45.519 --> 00:34:49.519
Map the model that does it the best right now, that's available right now.
00:34:49.679 --> 00:34:56.960
And don't just say we're gonna have a frontier model and we're gonna pay whatever price and we're gonna run all of our workflows through that.
00:34:57.119 --> 00:35:08.880
Because not only is that not sustainable, but if you do that, if you put too many of your workflows in one bucket that may disappear, so too may your company's progress overnight.
00:35:08.960 --> 00:35:10.559
And you don't want that to happen.
00:35:10.800 --> 00:35:14.159
So, yes, you need to build those model fallbacks and routing.
00:35:14.320 --> 00:35:20.000
So one gated tool can't halt your operations, but you gotta pay closer attention than ever.
00:35:20.159 --> 00:35:25.119
Uh, the fine prints of these model releases, uh, the restrictions, everything that's happening.
00:35:25.199 --> 00:35:28.960
And that's what we're gonna continue to do every single day on everyday AI.
00:35:29.119 --> 00:35:38.400
So I know this one went in slightly different direction, but that's the reality because we don't have a model to play with when it comes to GBT 5.6 right now.
00:35:38.559 --> 00:35:42.880
We can't do a hands-on anymore uh with Fable 5 either.
00:35:43.039 --> 00:35:51.119
Uh, so I think it's important to not just know what's new, uh, but why the US government is blocking it and what your business should be doing next.
00:35:51.199 --> 00:35:54.480
So I hope that we covered at least most of that on today's show.
00:35:54.559 --> 00:35:55.920
And I hope this was helpful.
00:35:56.079 --> 00:36:06.480
And as we do get access, you better believe whether it's uh GPT 5.6 soul or mythos, we're gonna continue to get more hands-on so your company can get ahead.
00:36:06.639 --> 00:36:10.239
So if this was helpful, please do me a favor and subscribe on the podcast.
00:36:10.320 --> 00:36:14.960
If you're listening on Spotify or Apple Podcasts, then make sure to go to your everydayai.com.
00:36:15.119 --> 00:36:22.400
We're going to be recapping the highlights of today's show as well as all of the other AI news you need to know to grow your company and career.
00:36:22.559 --> 00:36:23.280
Thanks for tuning in.
00:36:23.360 --> 00:36:26.159
Hope to see you back tomorrow and every day for more everyday AI.
00:36:26.239 --> 00:36:27.280
Thanks, y'all.
