Modern Creator
Matt Wolfe · YouTube

AI News: This Is The Make or Break Week

Anthropic's record IPO starts a months-long test of the AI bubble, three new models chase cheaper intelligence, and agents get a credit card.

Posted
2 days ago
Duration
Format
Talking Head
educational
Views
99.2K
1.6K likes
Big Idea

The argument in one line.

Anthropic's IPO turns AI bubble talk into something you can watch: whether insiders sell or hold as their shares unlock, while the week's new models show the race has shifted from smarter to cheaper.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • You build with AI APIs and want to know which of this week's new models gives the most capability per dollar.
  • You follow the AI market and want a concrete signal for judging whether the boom is a bubble, beyond headlines.
  • You pick a coding model for Claude Code, Codex or Antigravity and want benchmark context plus real hands-on tests.
  • You only catch one AI news recap a week and want the launches that matter in about half an hour.
SKIP IF…
  • You want deep financial analysis of the Anthropic IPO; the market section is a framing argument, not a valuation model.
  • You need rigorous model evaluations; several tests here are single runs judged by eye.
TL;DR

The full version, fast.

Anthropic raised $81.3 billion in the largest IPO ever and finished day one up 17%, but the argument here is that the first-day price is not the story. Anthropic's insider shares unlock in stages starting mid-December, so the next few months show whether employees and early investors cash out or hold, which is a direct read on how much the best-informed people believe in AI. The new models share one theme: nearly frontier intelligence for much less money. GPT 6.2 Astra is cheaper per task than Fable 5.2, Gemini 4 Neon sits one point behind Astra at under half the cost and tops a coding agent index, and Microsoft's MAI Code 2 claims 98% of GPT 6.1 Sol at 40% of the cost. Agents can now also pay for things through ChatGPT Wallet.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →
Chapters

Where the time goes.

00:00 – 00:26

01 · Cold open

Anthropic's record IPO framed as the first real AI bubble test, plus a preview of three new models, a one-sentence website builder and ChatGPT holding your credit card.

00:26 – 05:36

02 · Anthropic's IPO and the lock-up test

The IPO numbers, the staged insider unlock schedule from mid-December through February, why insider selling or holding is the signal to watch, quarterly disclosure, and what it means for OpenAI's 2027 listing.

05:36 – 10:36

03 · GPT 6.2 Astra and Sol

OpenAI's new top and everyday models pitched on token efficiency rather than raw intelligence, with Artificial Analysis cost per task, SVG benchmark results and a Megabonk game test.

10:36 – 11:37

04 · Sponsor: HyperAgent

Sponsored segment on a multi-agent team platform that added iMessage support and a free plan.

11:37 – 16:35

05 · Gemini 4 Neon

Google's mid-tier model priced at $3 and $15 per million tokens, scoring near the top of Artificial Analysis and first on its coding agent index, tested on SVGs and a Megabonk clone in Antigravity.

16:35 – 21:55

06 · The impress-me website test

Each new model gets the same open-ended prompt to build an impressive website. Astra and Sol produce near-identical deep-sea sites, while Neon builds a luxury watch site with a working loupe hover.

21:55 – 24:20

07 · Microsoft Ignite: MAI Code 2 and more

Microsoft's cheap coding model, its image editing model and a one-cent-per-minute realtime voice model, all pitched as close to the best at a fraction of the price.

24:20 – 26:05

08 · Gemini Notebook builds websites

The tool formerly called NotebookLM can now turn your sources into an interactive, shareable website, demoed by turning the week's news into a newspaper-style page.

26:05 – 32:00

09 · Rapid fire

HeyGen Avatar 6, Suno Visuals, Gemini computer use on Mac, Claude Duet for waitlist users, Claude Code desktop themes, Google Stitch Loop, ChatGPT Wallet, live dictation, ChatGPT for Sales and Genie 4.

32:00 – 33:00

10 · Wrap and sign-off

Recap of the three themes: the IPO as a bubble test, cheaper near-frontier models, and agents that can pay. No news video next week for Thanksgiving.

Atomic Insights

Lines worth screenshotting.

  • Anthropic's IPO raised $81.3 billion at $83 a share, passing SpaceX's June offering as the largest IPO ever.
  • The real AI bubble test is not the IPO price but whether insiders sell when their locked shares become sellable.
  • Anthropic let employees sell up to 10% of their stock on day one instead of the usual six-month lock-up.
  • By early February up to 1.6 billion Anthropic insider shares can be sold, more than the IPO itself offered.
  • If venture funds dump AI shares at the first chance, it signals they wanted a quick return, not a long-term bet on AI.
  • Anthropic's filing showed $16.3 billion in first-half revenue against $8.9 billion spent on compute.
  • GPT 6.2 Astra scores three points below Fable 5.2 on Artificial Analysis but costs about $4 per task versus $26.54.
  • The new GPT 6.2 Sol scored lower on a visual SVG benchmark than the older 6.1 Sol it replaced.
  • Gemini 4 Neon scores one point behind OpenAI's best model at less than half the cost and beats Google's own bigger Carbon model.
  • Gemini 4 Neon inside Antigravity ranked first on Artificial Analysis's coding agent index, ahead of Fable 5.2 in Claude Code.
  • Given the same open-ended prompt, GPT 6.2 Astra and Sol built nearly identical deep-sea websites with the same wording.
  • Models that regress to the mean on creative tasks produce the same theme and names even across completely different prompts.
  • Microsoft claims MAI Code 2 delivers 96% of Claude Opus 5.5's coding performance at 5% of its cost.
  • Every major release this week sold the same promise: almost the best model for a fraction of the price.
  • ChatGPT Wallet gives agents a stand-in card number, so the agent never sees the real credit card.
  • Claude Duet runs in the background and counts against normal Claude limits, so turning it on burns usage faster.
Takeaway

Watch insiders sell, and buy intelligence cheaper

WHAT TO LEARN

Anthropic's IPO gives you a dated, observable test of AI bubble fears, while this week's models show the best value now sits just below the frontier.

02Anthropic's IPO and the lock-up test
  • An IPO's first-day pop says little; insider behavior when lock-ups expire tells you what the best-informed people believe.
  • Anthropic's unlocks run from mid-December to early February, so heavy selling then would be an early warning for the whole AI market.
  • Public listing forces quarterly disclosure, so revenue and compute spend stop being rumors and become numbers you can track.
03GPT 6.2 Astra and Sol
  • Judge models by cost per task, not price per token, because a more token-efficient model can be far cheaper at the same rate.
  • GPT 6.2 Astra lands within a few points of Fable 5.2 at a fraction of the cost per task, which matters most for heavy API use.
  • A newer model can regress on specific tasks, as Sol did on SVGs, so rerun your own key tests before switching.
05Gemini 4 Neon
  • Gemini 4 Neon is the value pick of the week, scoring near the top overall and first on a coding agent index for about half Astra's cost.
  • Mid-tier models now often match or beat their own family's flagship, so default to the workhorse tier and upgrade only when it fails.
06The impress-me website test
  • Give models an open creative prompt to check taste; two OpenAI models produced the same deep-sea site with the same wording.
  • Repeated themes across unrelated prompts are a sign a model falls back on generic defaults, which shows up as same-looking vibe-coded sites.
07Microsoft Ignite: MAI Code 2 and more
  • Vendor claims like 98% of the performance at 40% of the cost come from the vendor's own tests, so treat them as a reason to test, not proof.
  • Cheap coding models like MAI Code 2 are positioned for simpler work, not long, complex projects, so match the model to the job.
08Gemini Notebook builds websites
  • Gemini Notebook can now turn your uploaded sources into a shareable interactive website from one plain-language request.
  • It is aimed at non-coders: dashboards, task trackers and study pages built from documents you already have, ready to share in a new tab.
09Rapid fire
  • Always-on agents like Claude Duet draw from your normal usage limits, so enabling them means hitting caps sooner.
  • Agent payments now use stand-in card numbers with per-purchase biometric approval, which limits exposure if you do hand an agent a card.
Glossary

Terms worth knowing.

Lock-up period
The window after an IPO when employees and early investors are not allowed to sell their shares, usually about six months, meant to stop a flood of selling right after a company lists.
Staged unlock
A lock-up that releases insider shares in batches on set dates instead of all at once, letting some selling happen early and the rest over the following months.
Artificial Analysis
An independent benchmarking site that scores AI models on a combined intelligence index and also tracks cost per task, speed and coding agent performance.
Cost per task
The total API spend a model needs to finish a benchmark task, which captures both token price and how many tokens the model burns, unlike a plain per-token price.
Token efficiency
How few tokens a model uses to complete the same work. A model with the same per-token price can still be cheaper overall if it is more token efficient.
Regress to the mean
Here, a model's tendency to fall back on the most statistically common answer, producing generic or repeated themes when given creative freedom.
Antigravity
Google's agentic coding app, comparable to Codex or Claude Code, which now offers some Claude models alongside Gemini.
World model
An AI model that generates an interactive 3D environment you can move through in real time, rather than a fixed image or video clip.
Resources

Things they pointed at.

00:26productAnthropic IPO (NASDAQ)
05:39productGPT 6.2 Astra and GPT 6.2 Sol
07:50toolBusey Bench (SVG benchmark)
09:10toolMegabonk clone test
10:36toolHyperAgent (sponsor)
11:42productGemini 4 Neon
14:55toolGoogle Antigravity
19:15channelTheo Browne on GPT 6.2 Astra
22:10productMicrosoft MAI Code 2
23:40productMAI Image 2.7 and MAI Realtime 1
24:22toolGemini Notebook (formerly NotebookLM)
26:20toolHeyGen Avatar 6
27:10toolSuno Visuals
27:38toolGemini computer use on Mac
28:20productClaude Duet
28:41toolClaude Code desktop custom themes
29:06toolGoogle Stitch Loop
29:45productChatGPT Wallet
30:45productChatGPT for Sales
31:10productGoogle DeepMind Genie 4
Quotables

Lines you could clip.

01:21
“I don't think that the first day price is really the story here. I think the story is what happens next because this IPO just started the first real test of whether or not AI is actually in a bubble or not.”
Clear contrarian take on the biggest story of the week→ TikTok hook↗ Tweet quote
03:30
“If the people closest to the company sell big at the very first chance they get, that tells you how much faith they have in where AI is heading as a whole.”
Standalone explanation of the bubble signal→ newsletter pull-quote↗ Tweet quote
19:19
“It regresses to the mean far too often.”
Short, quotable critique of GPT 6.2 Astra with a visual payoff→ IG reel cold open↗ Tweet quote
19:55
“Jellyfish, jellyfish, below the light, begin descent, begin descent, like even some of the same wording. That's crazy to me.”
Funny reveal that two models built the same site→ TikTok hook↗ Tweet quote
32:04
“Pretty much all the new models this week had kind of the same theme, which was you get more smarts for less money.”
One-line summary of the week's model releases→ newsletter pull-quote↗ Tweet quote
The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphoranalogy
This week, Anthropic went public. It raised $81 billion, making it the biggest IPO in history. And I also think it just started the first real test of whether or not AI is, in fact, actually a bubble.
Also this week, we got three new AI models that I put through all of my tests. We got a Google tool that builds you a website from just one sentence and ChatGPT wants to hold onto your credit card for you. All that and more in this week's AI news breakdown.
So let's get right into it. So here's what happened. On Wednesday night, Anthropic priced its IPO at $83 per share and they offered 980 million shares.
During that initial public offering, they raised $81 .3 billion. That actually passes SpaceX from back in June, which at the time was the largest IPO ever. So this IPO from Anthropic is now the largest IPO to ever happen.
Thursday morning, it started trading on the NASDAQ under the ticker. And when it opened, it opened at $111 .50 per share. It eventually closed at the end of the day at $97 .17 per share.
So essentially it finished up 17 % on day one. And as of this recording, the company's valued at roughly $1 .27 trillion. currently sitting at the 12th largest company on the planet.
Now, personally, I don't think that the first day price is really the story here. I think the story is what happens next because this IPO just started the first real test of whether or not AI is actually in a bubble or not. So here's what I mean by that.
Normally when a company goes public, the insiders are locked up. The employees and the early investors can't typically sell a single share for about six months. However, Anthropic did this a bit differently.
Their shares actually unlock in stages. On day one, employees were allowed to sell up to 10 % of their stock. That's up to 315 million shares.
The next batch unlocks in mid -December, about four weeks from the time I'm recording this video. And then there are three more unlock periods between January 5th and February 2nd. So by that point, up to 1 .6 billion insider shares can now be sold, which is well over the number of shares sold in the IPO itself.
And then most of the rest unlocks within six months with really only the co -founders and the biggest shareholders having to wait up to a year before they can finally sell some of their IPO shares. So going back to that whole comment about. This will help us determine whether we're really in a bubble or not.
Think about who's holding on to these shares. These are the people that know the company better than anybody else. The employees who built it and the investors who have seen every single number for years now.
And so over the next several months, every one of those people gets to make a choice. Do I cash out or do I continue to hold? Now, to be totally fair here, some selling is totally normal.
Employees want to go buy a house and venture funds need to go and return some of the money to their investors. So some selling shouldn't be like anything worth panicking over. But what I'm watching is how much people are unloading.
If the people closest to the company sell big at the very first chance they get, that tells you how much faith they have in where AI is heading as a whole. And then that kind of doubt can potentially spread to the whole AI market pretty fast. However, if they mostly hold when they could be walking away with billions, well, that's the people with the most information saying they think we're still early.
So mid -December here is really the first date I'm paying attention to. And, you know, I'll keep you posted with what happens around that. Now, two other things change now that Anthropic is public.
First, we actually get to see real numbers. Like a public company has to open its books every three months. Their filing already showed a company that brought in $16 .3 billion in the first half of the year, up from about $1 .2 billion a year earlier, while spending $8 .9 billion on compute.
So from now on, we get to see that every single quarter, and we don't have to rely on leaks or anything like that. Second, there's new pressure. There's a stock price now and shareholders want growth.
And the company that's watching all of this, the closest is probably OpenAI. Like their CFO told staff that they plan to go public in 2027 or sooner. I mean, we're at the end of November now.
So, you know, it's not going to be this year. But however, this plays out. the price and like the insider selling and all of that stuff, it's kind of setting the stage for what open AI is going to do.
Now, my hope is that when all of these unlocks happen, we see people holding onto their shares because if they're holding onto their shares and not immediately liquidating as much as possible, well then it's telling us we've got faith in the bigger picture of AI. However, if we start seeing some of these venture capitalist firms that invested in these AI companies, all start dumping their shares as quickly as they're allowed to.
Well, that shows these VC firms, they were trying to get a quick buck and they're not really in it for the long term and may not have as much faith in AI as they've led people on to believe, which, you know, could be the start of a bubbly, bursty kind of situation. All right, now let's talk about some new models. We got three big ones this week and I went and tested all of them.
First up, OpenAI released GPT 6 .2 Astra and GPT 6 .2 Sol. Now, I made a whole video about these. You can find it over on this video.
OpenAI GPT 6 .2 Astra is here. It's not what you think, but you know me. I'm going to give you the TL didn't watch in this video for those of you who only watch my Friday videos.
So Astra is OpenAI's top model, the one for like your hardest problems. Now, Sol is their cheaper, more everyday model. Both are live inside of ChatGPT work and in Codex if you're on one of the paid plans.
Now the pitch on this one isn't really that they're way smarter. Like they got a little smarter in some areas, but the pitch on this one is really that they're getting the job done using fewer tokens. They're like more efficient models and that the same work costs less.
Like on Frontier Code here, for instance. We can see 6 .2 Astra does slightly better than 6 Astra, but not by like a huge margin. However, when we look at 6 .2 Sol, it does almost as good as 6 .2 Astra and better than 6 Astra for quite a bit cheaper.
If you're using these models in the API, it actually got 20 % cheaper. So $8 per million input tokens and $40 per million on output for the Astra model. Well, 6 .2 Sol is at $2 for input and $10 per output, which 6 .2 Sol is actually the same price, but supposedly more token efficient.
So the net result is cheaper. Now on artificial analysis, GPT 6 .2 Astra is the second smartest model behind Claude Fable 5 .2, according to their benchmarking. However, if we come down here and look at cost per task, GPT 6 .2 Astra cost $4 per task at max effort, where Fable 5 .2 cost $26 .54.
So only a three point difference here on their benchmark, but quite a big difference in efficiency and how much it's going to end up costing you to run these models. Now on Busey Bench, Astra scored a 7 .4, putting it into sixth place on this benchmark. It cost a buck 42 and took about five and a half minutes to run.
6 .2 Sol scored a 6 .9, dropping it into 18th place, but it did it in about 68 seconds and for 14 cents. And one thing that I find very interesting here is that the older 6 .1 Sol is still sitting in number one on the leaderboard with a score of eight overall. So on this specific test, the new Sol actually took a step backward from the model that it replaced.
But again, this is also a visual benchmark. So, you know, use your own eyes and decide whether you think it's better or not. I also ran the pro versions through it.
Asta pro got a 7 .6 and it cost $4 and 24 cents and took almost 17 minutes to run. So pro got a 6 .8 to almost five minutes and costs about a buck at 4 cents. So if you're wondering which models are better at drawing SVGs, I don't think.
Pro models are necessarily worth the upgrade because the regular models did better. All right, now on my Megabonk test, Astra built probably the best opening screen I've seen from any of these models. Probably because it generated these scenes from the GPT image model, if I'm guessing.
When I start a run here, the game plays exactly how it should. I mean, my skateboard is facing the wrong way and these level ups are happening way too fast. But other than those two things, it looks pretty good.
And it plays pretty good. I'd actually put this ahead of what Fable 5 .2 did, but maybe close to what Opus 5 .5 is capable of. Now, quick side note, I only ran the Megabonk test on Astra.
I didn't actually run it on Soul. So bottom line on this one, I do think Astra is a pretty solid model that's cheaper to run and really good at visual work like 3D design and games. Now, the people that tested it early found that it stops too soon and could be hard to follow on big projects.
And OpenAI's own report says it's more likely to leave out problems it ran into along the way. I think if you're happy with something like Claude, I don't really see a reason to switch. But if you're a ChatGPT user, you already have it.
So, you know, go try it on something visual. Now, if you want a deeper dive, including why OpenAI says this is their first public model to cross like a new line on AI helping to build AI, that's in my other video from earlier this week. I also break down what everybody says they hate about these new models.
So I'll make sure that video is linked up below. And also I've got one more test that I'm going to run on these two models that wasn't in that video. So I'll show that in just a minute.
The next update is from HyperAgent. It's a platform where you can build a whole team of AI agents that work together, kind of like... a team of virtual employees.
And they just added iMessage support. So I could just text my agent to work on something if I'm away from my computer. So for my show me your workflow videos, I've got it set up so that one agent researches the person I'm interviewing and their background.
Another comes up with interesting questions based on that research. And then a third turns everything into an interview outline. And they all work together in this virtual room alongside me and my human team.
So they like give each other feedback and then we can jump in and direct them to. and HyperAgent just launched a free plan if you wanna try it out. Or if you're already paying for ChatGPT, you can sign in with your existing subscription and use supported OpenAI models inside HyperAgent without paying extra for the model usage.
So check it out at the link in the description box and thanks so much to HyperAgent for supporting my channel and sponsoring this portion of today's video. All right, so back to the new models. OpenAI wasn't the only one shipping this week and I think this next one might actually be a better deal.
This week, Google released Gemini 4 Neon. Now, if you're keeping track of the Gemini 4 family, you have Carbon, which is the big one for like the hardest, longest projects. Think of that one as like the equivalent to Fable from Anthropic or Astra from OpenAI.
Hydrogen is the smaller, faster model. If you're looking at the Anthropic equivalent, it's probably comparable to like Sonnet. And if you're looking at OpenAI's models, I would probably put it in line with like their Luna model, I guess.
Now this new model, Neon, is the one that sits right in the middle. So like Anthropix Opus and OpenAI Sol models, probably like the workhorse model for most of these platforms is the one that this is in the tier of. Now this model is priced at $3 per million input tokens and $15 per million output tokens with a million token context window.
So comparing it to Opus, slightly cheaper. Opus is $4 and $20 versus Neon's 3 and 15, but slightly more than GPT 6 .2 Sol. But right around that pricing tier as well.
You can use it inside of the Gemini app if you're an AI pro or ultra subscriber. So it's not. on a free tier, but if you're on one of the paid tiers, you should have access to it.
And of course it's available inside the API. Now, here's why I'm paying attention to this one. And I actually think this is a really good model and a really good deal.
On artificial analysis here, Neon scores 61. So just one point behind 6 .2 Astra and four points behind Fable 5 .2. Now, keep in mind that OpenAI's GPT 6 .2 Astra is their best model at the moment.
This is their state -of -the -art frontier model. And it's only one point more intelligent than Neon at like... less than half the cost.
And Google themselves says it's 32 % lower cost per task than carbon, which by the way, this new neon model, which is 32 % less than carbon is scoring one point higher than carbon. And if you're building apps with AI, this is something to look at right here on artificial analysis, coding agent index. Google Neon is number one with a score of 59.
Like that's beating out Fable 5 .2 inside of Claude Code, and that's beating out 6 .2 Astra inside of Codex. So the combination of Gemini 4 Neon and anti -gravity is winning this artificial analysis coding agent index. So let's see how it did on my tests.
On Busey Bench, Neon scored a 7 .1. That puts it at 12th place on the leaderboard. And it did it for 70 cents and took about five minutes to generate this SVG.
So that's about half the cost of Astra for a score that's three tenths lower. Now, carbon still showing up in second place. So carbon still seems to be a better model at generating SVGs.
But I mean, it's the difference between a 7 .9 and a 7 .1. So not a huge, huge leap. But like I always say, use your eyes.
Don't pay attention to these numbers. you can decide which one you think does the best at drawing. Now, I did fire up anti -gravity to tell it to create a clone of the popular 3D game Megabonk.
And one thing I do need to mention is that anti -gravity has gotten a lot better than it used to be. I thoroughly enjoyed using anti -gravity now. It feels on par with like Codex.
Maybe I won't go that far. I still like Codex a little bit better, but anti -gravity is catching up. And one thing that's also cool about anti -gravity, is if you look at the models that are available it's not only gemini models now they actually have access to some of the clod models available inside of anti -gravity so let's see how it did on the megabonk test i actually am testing this for the first time right now so on this main home screen we can see what the different characters look like sir ufi fox the fox looks decent dice head clank and calcium so calcium that's the skateboarding skeleton that's what not supposed to look like let's go ahead and use this one at least the skateboard's facing the right way this time all right so camera motion works i can move around the camera uh we got some bad guys coming towards us i could jump i could power slide by hitting shift let's see here let's up level to our six shooter here i mean mechanically it's playing well the graphics aren't the most amazing graphics i've ever seen from one of these uh vibe coded games but
It looks decent. It looks decent. I definitely think that Astra did better with the graphics, but mechanically everything's nailing it now.
Like pretty much this prompt, it always nails the gameplay every single time. The only difference is how decent the graphics look at this point. And you know, they're okay.
Not the best graphics we've seen so far. All right, now that other new test that I mentioned, I've actually done this test in the past and then kind of stopped and went, why did I stop doing that? So I'm bringing this one back.
I've started giving every model this exact same prompt just to see what it does. Build me a website that will truly impress me. Use your capabilities to create something that's truly spectacular and visually appealing.
The subject matter of the website is up to you. Just impress me. I want to see what it comes up with when it gets to choose.
There's a reason I'm starting to do this test a little bit more. I said on my video that I put out on Tuesday that vibe coded websites are all starting to kind of look the same to me. So I want to know which of these models have some taste when nobody's telling it what to do or specifically steering the design.
So I ran this same prompt on the three models we got this week, Astro Soul and Neon. So this is my first time actually seeing these. I haven't reviewed them yet.
This first one here is what Astra generated from OpenAI. This site is called Paylagic Ocean Atlas. We've got some images here.
This one's actually like a slightly animated jellyfish. Then we've got like some like kelp beds or something. And then...
a squid and there is some like minor minor animations going on let's see begin descent all right so if we change this it changes the top here all right so let's go way deep and log this depth i don't know what that means save entry yeah i don't know what this little thing is doing but okay so you've got the sunlight zone okay so this is the whole page this is what astra built for me that's it about the atlas okay yeah that's it that's the whole site if there's sound i guess cool all right that's what astra 6 .2 made next up let's see what opus 6 .2 made okay so this one says hodl below the light Leave the surface behind.
Follow the last of the sunlight. Begin the descent. And then we've got depth zones.
Yes, I gave the same prompt to two different models. I wasn't expecting it to essentially try to make the exact same thing. It literally made like the same site just twice with some slight variation.
Okay. And it's not even that impressive. Yeah, there's some animations.
Let's see. Auto dive. Cool.
We can see the depth as it scrolls down, but I mean, nothing insanely impressive. Ping sonar. Yeah.
Okay. Sound off. Let's turn sound on.
It's the same sound. So I saw this thing from Theo Brown the other day where he literally wrote up, here's why I hate GPT 6 .2 Astra. And he wrote all of this stuff about it, but there's one section that I highlighted here that really stands out to me.
It regresses to the mean far too often. Ben and I, his podcast co -host, both had it make a unique level in different video games. We both got water plus music plus bell themed levels named Bellwether.
Entirely different games, prompts, setups, etc. And we both got the same name and the same theme somehow. I feel like I just got the same thing when I told it to impress me.
Astra gave us this undersea themed website that lets us explore different depths, but... isn't really that interactive. And Seoul gave us this website that lets us sort of explore the depths with the deep sea theme that isn't really that interactive either.
But they definitely fell into like the same exact theme, right? Jellyfish, jellyfish, below the light, begin descent, begin descent, like even some of the same wording. That's crazy to me.
All right, now let's check out what Neon did. Okay, this one went in a totally different direction. mason voucher when i move it around you can see the area light up behind it you've got this like analog watch kind of thing with mechanical gears assembled skeleton exploded interesting okay so you can sort of change the view of this watch the exploded view doesn't look that much more impressive than the regular view let's see okay You can stop that now.
Okay, so that makes songs. Inspect the trilogy. All right, inspect under the watchmaker's loop.
Oh, that's really cool. The little magnifying glass that it generated. Look at that.
That looks really cool, actually. So, I mean, Neon did a better job, I think, than both Sol and Astron doing something that would impress me. I really like the hover.
I know that's probably, if you're a coder, that's like a really easy feature to implement, but I'm easily impressed, I guess. So that's Google's neon model, which in my opinion is probably like the best bang for your buck right now, especially if you're going through the API. It's a really decent coding model.
It's inexpensive to use. It's up there on artificial analysis is one of the best models. It makes impressive looking designs.
Like for the longest time, we were all talking about Google dropping the ball and Anthropic and OpenAI passing them while Google was quiet. I feel like with the models they're putting out now, like. they're kind of caught up again.
I don't think we can ever count out Google. Like they have the most training data available to them out of all of the companies. And you know, I think, I think they cooked with this neon model.
I think this is a really good model that I could see using as more of a daily driver quite often. All right. One more model real quick.
And this one is mostly focused on price. So Microsoft held their ignite conference this week. And, you know, to be honest, most of what they announced was for like big enterprise type companies.
But there was a few things in there for us who actually build with AI at a much smaller level, like this new model called MAI Code 2. And this is Microsoft's own coding model. So Microsoft says that this model delivers 98 % of GPT 6 .1 Sol's performance at 40 % of its costs.
Now, this came out. like the same day that OpenAI launched 6 .2, which, you know, OpenAI has a tendency to do that kind of stuff where they'll launch things on the same day that other companies are doing keynotes to try to like overshadow them. Although they claim they don't do that, but they definitely do.
But anyway, they also claim that it gets 96 % of what Cloud Opus 5 .5 can do at 5 % of its cost. They claim that with this model in blind front -end app evaluations, developers preferred its results over GPT -6 Sol 64 % of the time. Now that is front -end apps.
So I'm guessing that's more like the design. People like the way the designs look. Now, all of these are numbers given to us by Microsoft.
Now, Microsoft also says that this one's not really tuned for the hard, really complex, long projects, but for simpler stuff, it is a. pretty cheap model to use for coding. It's in GitHub co -pilot right now.
And Microsoft does say that it's coming soon to open router and Vercel. And as soon as I'm able to run it, I'm going to get it through Busey bench and do all my normal tests. Now, as for the rest of ignite, just like real quick, Microsoft also shipped a new version of its image model, MAI image 2 .7 and on artificial analysis, it actually sits at the number one model when it comes to image.
editing. Now, when it comes to text to image and generating from scratch, it sits in fourth place behind all the chat GPT image models. They also released their own real time voice model called MAI real time one, which starts at just one cent per audio minute.
And pretty much the pitches on like all of this stuff they promoted at ignite. They're all kind of the same thing. They're close to the best for a fraction of the price.
So not quite as good as the best of the best, but a lot cheaper than the best of the best. And like, Almost there.
Okay, so for this next thing, if you've ever wanted to build like a little website or a tool, but you don't code, you're gonna like this. Google added a new feature to Gemini Notebook this week. If you don't know what Gemini Notebook is, it used to be called Notebook LM.
It's the tool where you load up your own notes and documents and research, and then you can ask questions about them. Now you can type, make me a site, describe what you want. and it builds an interactive website out of your sources.
So Google's examples are things like dashboards for your quarterly sales numbers, a tracker for tasks and deadlines, and, you know, like interactive websites to study for exams and things like that. And then when it's done, you can open it in a new tab and share it with somebody. Now, this one's rolling out right now this week, and I haven't tried it yet, but I do have this notebook inside of a Gemini notebook where I loaded in a bunch of the news from this week.
It wasn't able to pull all of it in because some of the sites blocked it. But let's go ahead and say, make me a website that breaks down all the news for this week. And about two minutes later, this is what I've got.
It generated this newspaper looking website for me that put all of the news in there. So the big headline news, Anthropic's $1 .8 trillion debut crowns a relentless week of frontier releases. I mean, that sounds about right.
And we can see it's got everything else in here that I talked about. Anthropic IPO, Gemini Neon, MAI Real Time, Gemini Notebook. GPT 6 .2, Microsoft Ignite.
All of it is in here and it looks really cool. I mean, it put charts in here. It put little like tables in it and it presented all of the news that I just pasted in there into a nice visual website layout.
I'm pretty impressed with what this does. And then down here, we have all of the sources that we can even filter. Look at this.
That's pretty sweet. All right, so those are all the big stories of the week, but there's still a few more things that I want to get into. So let's knock it all out in a rapid fire.
HeyGen released Avatar 6 this week, which is their newest model for making a talking video of yourself. You give it one photo and there's no set limit on how long the video can be. So some of their examples run 25 minutes or longer with a single continuous take at 2K resolution.
They claim it gets an 84 % on the Turing test. So they showed people a real clip and an AI clip of the same person saying the same thing and asked which one was real. And people picked the AI clip 42 % of the time.
Now, if that was 50%, that means like it was a 50 -50 shot that nobody could tell the difference. So it's not quite fooling everybody yet, but the examples I've seen look really good. So glad you stopped by.
If you are picking up a prescription, please head to the counter on the left and have your name and date of birth ready. I haven't experimented with this one yet myself, but it's definitely something I plan on playing with in a future video. Suno, the AI music tool, launched visuals.
So you drag a song from your library into the new visuals tab and it makes lyric videos, album art, and video clips to go with it. The clips follow the structure of the song and could lip sync to the lyrics. And this is available on web, iPhone and Android.
Now the clips do max out at 15 seconds each and visual credits are sold separately from your song credits. Google's Gemini app on the Mac can now use your computer for you. You describe a task and it clicks, types, and fills in forms across your apps and your files.
It works in its own background window so you can keep your screen and your mouse while it runs. You pick which apps it can touch and it asks before any big action. Here's an example where they give it a drawing of a butterfly, give it access to computer use, and then once it starts using the computer, it recreates that butterfly inside of Blender and then creates an animation, all just taking control of your computer for you.
Right now it's US only and you need a map with Apple Silicon, and it's only for the paid tier of Google AI plans, not something I've actually tested yet. We got a quick update on Claude Duet, Anthropix always on assistant.
I covered this one a few weeks ago, but now everybody in the US on a pro max or ultra plan who was on the wait list is now able to use Duet. One thing to keep in mind is that it runs in the background and counts against your normal Claude limits. So you will go through your Claude.
limit faster. Sticking with Claude here for a second, Claude code on desktop now lets you build custom themes. You tell Claude how you want the app to look like the colors and animated backgrounds and whatever you want.
And it rebuilds its own interface to match. Now it only works in local sessions in the desktop app for now. And you know, it's not going to add any sort of productivity or be that useful, but it's fun, I guess.
Google Labs announced Stitch Loop, and if you've ever vibe coded apps before, you're probably gonna dig this one. So pretty much anybody could vibe code apps these days, but like most vibe coders aren't really thinking much about security and the speed of their apps. So Stitch Loop connects to your code and to your app's data and keeps on running checks for things like security and speed and accessibility.
When it does find a problem, it builds the fix and sends it to you to approve. So it's not going to do anything unless you say yes. Now, right now, this one's on waitlist only, and I haven't tested it yet, and we don't know what it's going to cost.
But if you do a lot of vibe coding, it's probably going to be very helpful, especially if you don't ever think about security yourself. Some news out of OpenAI, we got ChatGPT Wallet. Now, they built this one with American Express, Chase, MasterCard, and Visa, and it lets your dot, you know, your always -on assistant, buy things for you.
So you add your credit card once, you approve every purchase with face ID or your fingerprint, and then the agent gets a stand -in card number. So it actually doesn't ever see your real card number. Right now it's US only and it's only credit cards.
You can't use a debit card. You also have to be on a pro plan, which is the plan that starts at a hundred bucks a month. I haven't set this up yet.
I'm still not super comfortable giving my agents credit cards. But if you are somebody that's tested this, let me know in the comments what you think so far. ChatGPT also got live dictation this week.
So when you talk to it, your words now show up as text while you're still speaking. So you're not sitting there waiting for it to transcribe. If you talk your prompts instead of typing them, this might be a nice new little upgrade.
And OpenAI also released ChatGPT for sales. It's a set of like ready -made workflows for salespeople. So things like building a list of prospects or prepping for a call or writing the follow -up or flagging deals that are at risk.
It'll plug into more than 30 different tools, things like HubSpot and Salesforce and Gong and Granola and DocuSign. And if you run a small sales team, this could be really helpful to you. The one downside, it's only on business and enterprise workspaces right now.
And finally, Google DeepMind released Genie 4, their newest world model. I have actually demoed this in a couple previous videos now, but you type a prompt and it generates a world that you can move around in in real time. With this version, you can type new prompts while you're inside of it to change the weather, add characters, or change what's happening, and it generates the sounds as you go, including things like characters talking.
Now, Genie 4 is only available for Google AI Ultra subscribers right now, and the pro subscribers can join a waitlist. However, they also bumped their older Genie 3 model up to three minutes per world, which used to only be one. So that's another update too.
And that's the week. The company behind Claude is now a public company, and we're going to see our first real test of whether or not we're in an AI bubble. Pretty much all the new models this week had kind of the same theme, which was you get more smarts for less money.
Not the smartest thing in the world, but cheaper and almost as smart. And AI agents can officially pay for things for you now, even if most of us, you know, me included, aren't quite ready for that. Quick heads up, next week is the week of Thanksgiving and I'm going on a family vacation, so there will not be a news video next week.
I don't think there's going to be any huge announcements. on the week of Thanksgiving here in the US. But if there is, I will make sure I break them down in a video when I get back from vacation of any news that we might have missed while I was gone.
Oh, also, if you want that full breakdown of GPT 6 .2, that video will be on the end screen at the end of this video, maybe even on the screen right now. But, you know, thank you so much for hanging out with me and nerding out with me today. And hopefully I'll see you in the next one.
Bye bye.
The Hook

The bait, then the rug-pull.

The title promises a make-or-break week, and the opening line names the stakes: Anthropic just went public, and the next few months of insider selling will show whether AI is a bubble. Everything after that is the week's evidence on the other question, which is how fast frontier intelligence is getting cheap.

Frameworks

Named ideas worth stealing.

02:41concept

The insider-unlock bubble test

  1. Watch how much insiders sell at each unlock date
  2. Some selling is normal (houses, fund returns)
  3. Heavy selling at the first chance signals doubt
  4. Mostly holding signals the best-informed think it's early

Use insider behavior at staged lock-up expirations as a read on whether the people with the most information believe in the company and the wider AI market.

Steal forjudging any hyped sector's IPO
11:42list

Three Gemini 4 tiers mapped to rivals

  1. Carbon: biggest, like Fable or Astra
  2. Neon: workhorse middle tier, like Opus or Sol
  3. Hydrogen: small and fast, like Sonnet or Luna

A quick map of where each Gemini 4 model sits against Anthropic and OpenAI's lineups.

Steal forexplaining model lineups to non-technical readers
16:37concept

The impress-me taste test

Give every model the same open-ended prompt to build a website that will impress you, with the subject left up to the model, to see which ones have taste when nobody steers the design.

Steal forcomparing creative output across models
CTA Breakdown

How they asked for the click.

VERBAL ASK
32:35next-video
“If you want that full breakdown of GPT 6.2, that video will be on the end screen at the end of this video.”

Soft end-screen pointer to the deeper GPT 6.2 video, set up twice earlier in the episode, plus a note that next week has no news video.

MENTIONED ON CAMERA
Storyboard

Visual structure at a glance.

open
hookopen00:00
IPO headline
promiseIPO headline00:50
unlock schedule
valueunlock schedule02:05
benchmarks
valuebenchmarks07:15
Megabonk test
valueMegabonk test09:05
Gemini 4 Neon
valueGemini 4 Neon11:42
deep-sea site
valuedeep-sea site17:56
watch site
valuewatch site20:56
MAI Code 2
valueMAI Code 222:45
Notebook site
valueNotebook site25:46
rapid fire
valuerapid fire27:42
sign-off
ctasign-off32:47
Frame Gallery

Visual moments.

One-click upgrade to your Google

Get more breakdowns in your search results

Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.

Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Watch next

More from this channel + related breakdowns.