← Back to search

Claude Fable 5: The best AI model in the world - just filtered

KI und Tech to Go - der Praxis-Pitch · 2026-06-11 · 105 min
relevance 55 13001 words Episode page ↗ Audio ↗
Show full episode description
Anthropic hat mit Claude Fable 5 das stärkste KI-Modell veröffentlicht, das man je kaufen konnte – und gleichzeitig entschieden, dass niemand es ganz bekommt. Die Öffentlichkeit erhält die Version mit Sicherheitsfiltern, das ungefilterte Schwestermodell Mythos 5 bleibt rund 150 ausgewählten Organisationen vorbehalten. Und das vier Tage, nachdem Anthropic selbst vor zu schneller KI gewarnt hat. Yusuf Sar ist diese Woche solo am Mikrofon und erklärt, warum an diesem Release gerade fast alles hängt, was die KI-Welt bewegt. Den Anfang macht Fable 5 selbst: Platz eins in den unabhängigen Benchmarks und ein eigener Nachttest, bei dem das Modell ein seit Wochen festgefahrenes Projekt im Schlaf gelöst hat – aber auch der doppelte Preis, der Rauswurf aus der Abo-Flatrate und eine Datenspeicher-Klausel, die für DSGVO-Verträge zum Problem wird. Danach räumt Yusuf die Apple-Keynote ab: Das neue Siri läuft auf einem Google-Gemini, und Apple baut sich in aller Ruhe eine Mautstation zwischen KI-Laboren und iPhone-Nutzern. Es folgt die harte Ökonomie: GitHub Copilot rechnet künftig pro Token ab, IT-Budgets werden unkalkulierbar, und selbst Google mietet für 920 Millionen Dollar im Monat Rechenleistung bei SpaceX. Den Gegentrend zeigen kleine Modelle wie ZAYA1-8B und der lernende Agent Hermes Desktop – gute KI kann auch lokal und günstig laufen. Im Deep Dive wird es dann ernst: Laut Anthropics eigenem Bericht schreibt Claude bereits über 80 Prozent des eigenen Codes, während dasselbe Mythos-Modell bei der NSA als Cyberwaffe im Einsatz ist. Und am Ende will sogar der Staat mitbesitzen – Bernie Sanders fordert die Hälfte der KI-Konzerne für die Allgemeinheit. Was passiert, wenn die beste Intelligenz der Welt gleichzeitig rationiert, verteuert und militarisiert wird? Die Antwort der Folge ist bodenständig: Datenschutz-Zusagen prüfen, Token-Kosten unter Kontrolle bringen – und jetzt anfangen, mit lokalen Modellen eigene Kompetenz aufzubauen, bevor die erste Schock-Rechnung kommt.
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Anthropic's Claude Fable 5 launch raises AI-inequality concerns via filtered public vs unfiltered partner versions.
Benefits
  • Strongest coding model with 80.3% on SIE Bench Pro
  • One million token context window
  • Up to twelve-hour autonomous agent runs
  • Filtered Fable 5 buyable by anyone via credit card
Use cases
  • Fable 5 scores 80.3% on SIE Bench Pro vs Opus 4.8 at ~69, OpenAI ~59, Gemini 3.1 Pro at 54
  • Fixed a weeks-long unsolved coding project overnight: a lead generation/tracking system consolidating Google Analytics, Ads, Search Console, inbox and CRM data
  • Simon Wilson burned $110,000 on first day testing, $99 for a single long agent session
  • Ranked #1 on Artificial Analysis Intelligence Index, ~5 points ahead of Opus 4.8
  • Ethan Mollig reports runs of up to twelve hours processing multi-page tasks unaided
KPIs / results
  • 80.3% on SIE Bench Pro (Opus 4.8 ~69)
  • 1 million token context (~100,000 words)
  • $10 per million input tokens, $50 per million output tokens
  • Mythos 5 expanded from 50 to 150 organizations across 15+ countries
Tools / build
  • Claude Fable 5
  • Claude Mythos 5
  • Project Gly Swing
  • Artificial Analysis Intelligence Index
  • Lead generation/tracking system (Google Analytics, Ads, Search Console, CRM)
0:00 / 0:00
🌐 This transcript was automatically translated to English from the original.
and welcome to a new episode of AI and Tech to Go's practical pitch. I'm Yusuf Sar and today I'm alone here at the microphone. Unfortunately, Christian can't be there today because he's on vacation. That means you get a solo version today, no discussions, no back and forth, just my notes. You can see there's quite a bit of material, there's even a real banger. Where do I start? Normally I would be pondering now, but this week the answer is clear. The story of the week just came in yesterday. I divided the whole thing into six stations. So firstly the main story about Tropic, the company Intercloud, released their new top model Fable 5 yesterday. And this isn't just another new release. Pretty much everything depends on that this week. Two class AI, IPO, data protection. That will be the big chunk today. Secondly, I'll briefly clear up the Apple Keynote afterwards. As we already announced, the new series runs on a Google Model. Oh wonder! But in the summer the show was significantly thinner. Yes, we'll do that compactly. Third, let's talk about money. So about the tough economics behind AI. Why your AI tool could suddenly become incalculably expensive. And fourthly, let's take a look at the new models outside of Fable. There are some really exciting things there. Particularly with the smaller models, they fluctuate. And fifth, and this is my deep dive today, things get a little weird. The point is that AI is now driving its own code and at the same time being used as a cyber weapon. And that's more closely related to Fable Five than you might like. And sixthly, finally the big picture. Politics discovers AI. So get dressed, grab a coffee or whatever you're drinking. And we get started. So let's start with the story that outshines everything else this week. Yesterday on June 9th, Anthropic, the company behind the AI ​​model Claude, which we use most of the time here, released its new top model. And that's called Claude Fable Five. And first of all, this is not just another new model, as I would have already said. This is a step. So that basically changes everything. Yes, technically, in terms of price. And also the question of who actually gets access to the best AI models. Just to classify it so that you understand where the thing belonged. Anthropic currently has three model classes. So sorted by size and price. Haiku, this is the small, fast, cheap model. Sonnet is the middle class, so to speak the workhorse. And Opus, the previous top model. Fable Five is now not a new version of one of them, but a completely new fourth tier, above Opus. So Anthropic calls this the Mythos class. And my dears, when you hear the word myth, it should ring true with you right now. That's the code name that's been floating around in reports for weeks. Anthropic's supposedly dangerous secret model, kept under wraps. But now comes the real twist to this announcement. Anthropic didn't introduce one model, but two. Cloud Fable Five, basically five for the public. And Cloud Mythos Five for selected partners. But these are not two different models. So Anthropic said that themselves. Fable and Mythos Five have identical weights. So weights are basically the billions of parameters that are the model of the brain. The only difference is the security layer all around. So basically the rail guards. Fable Five is the version with a thick protective filter that anyone can buy. So basically us too. And Mythos Five is exactly the same model without these light planks. And only the chosen ones get that. Even the name, the names are a little play on words. Anthropic explains this itself. Fable comes from Latin. Fabula. That's what we tell. So that means what is being told. And in terms of linguistic history, that is the same word as the Greek myth. So same model, two narratives. Someone already thought of something with the name. But give it a spin. First of all, how good is this thing anyway? And the short answer is, it is by far the strongest AI model you can currently use. I'll give you the most important numbers. So the most important benchmark for programmers is called the SIE Bench Pro. This is a test in which the model has to fix real bugs in real large software projects. So no textbook assignments, just real work on real code. Fable 5 achieves 80.3%. The previous top model Opus 4.8 is at a good 69. The best OpenAI model at just under 59. Google Gemini 3.1 Pro at 54. That is no longer a narrow lead. This is basically a completely new league. And in a second test called 40 Code, which sets significantly more difficult production-level tasks, Fable 5 manages more than twice as much as the old opus. And lest you think, these are just Antropic's own glossy films. Artificial Analysis, which is an independent service that systematically tests all major models and makes them comparable, immediately placed Fable 5 at number 1 in its Intelligence Index. Around five points ahead of the best model. By the way, that was Opus 4.8. Five points doesn't sound like much, but on the scale it's quite a huge difference. What does that mean in practice? I find three things remarkable. First, the context window, a million tokens. Context window means this is how much the model can keep in mind at the same time. A million tokens, that's roughly just 100,000 words. So several thick books or a complete company codebase. Secondly, endurance. Ethan Mollig, a professor at Wharton Business School who is one of the first to seriously test every new model, reports runs of up to twelve hours at a time in which the model processes a multi-page task description completely on its own. Twelve hours, that's basically a digital employee who you give a task in the morning and it's done in the evening. And thirdly, the verdict from Simon Wilson, one of the most famous independent developers on the scene, who has been dissecting every model for years and really doesn't follow anyone's words. After his first day of testing he wrote that the model was, quote, a beast. So a beast. Slow, expensive, but it practically eats through everything. And in a single session he delivered work that would have taken him several days to complete, including tests and documentation. And of course I tested it myself. I started the Fable yesterday night and because of one of my most complex coding projects, which I haven't been able to solve for weeks, even with Opus 4.8, I simply said overnight, just fucking fix it. And this morning I got up and it's fixed, it's running, it's clean and it works. So and it's really not a simple product. So that was specifically, I built a kind of lead generation, lead tracking system, where I consolidate data from Google Analytics, Google Ads, Google Search Console, from our inbox, from our CRM system and throw all this data together and then try to canonically understand which customer came via which channel, via which click, via which campaign exactly, how much sales they made and what I have to do so that I can repeat that. Well, it really sounds simple, but it's a huge project. Anyway, now comes the care page and it also has a lot to offer, namely the money. Fable 5 costs $10 per million incoming tokens and $50 per outgoing token via the API, i.e. the programming interface through which companies and developers integrate the model. That's actually, to be precise, exactly twice as expensive as the previous Opus 4.8. I would have imagined it to be more expensive, but it's still expensive. And in the normal cloud subscription, i.e. in the plans from $20 to the Max Account, which I also use, Fable is only included until June 22nd. So it's a kind of trial phase and that counts twice towards the usage quota. From June 23rd it will be removed from the subscription and will cost extra via so-called usage credits. So these are these consumption credits that they have now tried to make us attractive, like prepaid cards. I think with the Max account it's somehow $200 or $200 that you get there. But it's actually a joke. This is actually a novelty. So for the first time, the best cloud model is permanently not included in the flat rate. So the best intelligence is only available at an additional cost, calculated based on consumption. So you can keep that in your head, this thought. And because we will talk about the end of the flat rate later in the money block. Yes. And to give you a sense of the scale, the Simon Wilson I was talking about earlier burned $110, so $110, burned $110,000 on his first day of testing. Yes. In a single day. Of which $99 is for a single long agent session. so cool, breathtaking abilities. But the counter also goes up in a breathtaking way to see how much it costs. And now another part that concerns me too. The thing about the two versions. So who gets this unfiltered Mythos 5? Of course, we've already discussed that. The selected organizations that we talked about last week from the program called Project Gly Swing. Originally there were only 50 organizations and now since the beginning of June there have been 150, so the organization has expanded to over 15 countries. A fundamental debate immediately broke out in the tech community about the dichotomy. On Hacker News, which is basically the most important discussion forum in the tech industry, one of the most discussed posts was titled: Fable 5 feels less like a product launch and more like a preview of AI inequality. And that's exactly the point. So two class AI, the chosen governments, hand-picked partners, get the full model and the rest of the world gets the throttled version. And I think this concern can't just be brushed aside now, because it could actually become a blueprint, yes, for top models, i.e. how they are distributed. To be fair, you have to explain how this throttling works in concrete terms, because that's not at all unintelligent. In front of the model are so-called classifiers, i.e. guard models that check every request before it reaches the actual model. And if a watchdog raises a question about cyber attacks or biological and chemical weapons, then Fable 5 doesn't get a request, but automatically the older Opus 4.8. So, according to Anthropic, this happens in less than 5% of sessions. It will be displayed to you and you will not pay the wrong price for this answer. So far so understandable. The problem is the false alarms. So on Hacker News a medical physicist reports accordingly. I use the word nuclear all the time professionally. So probably a radiologist. And just like that, he throttled back. Another wanted to have medical image data evaluated. A completely normal research task. And was immediately flagged as a bioterrorism risk. So even a harmless question about malaria was blocked. And an ordinary security audit of a code was classified as a cyber risk. In plain language, this means the people, medicine, research and IT security, i.e. the serious specialist users, who end up with bad models more often than average, without meaning any harm. If you work in an area like this, be sure to test it with your own use cases and check it out. And then there is also a third filter level. And it's not quite as funny anymore. It came out, Anthropic also confirmed it. Fable 5 recognizes requests related to the development of competitive, cutting-edge AI models. So if someone wants help building a large model themselves, all this distilling, and for such requests the model intentionally delivers worse answers, without any hint, without any error message. It's just secretly getting dumber. Anthropic says this only affects 0.03 percent of all traffic. Maybe so, but the point is a fundamental one. And Nathan Lambert, who is one of the most respected AI researchers at the Allen Institute, not one who reflexively shoots against Anthropic, put it to him, he basically says, an AI model that automatically becomes less intelligent without informing me about it is essentially exactly the kind and a very complex AI that security researchers always warn us about. And it is also questionable that it was built this way on purpose this time. And now you have to add one more detail, then maybe it will become a shoe. It was only in February that Anthropic accused Chinese competitors of carrying out so-called distillation through thousands of fake accounts. So distillation, in German, distillation means that you let the competition's top model generate answers en masse and use these answers to train your own cheaper model. You basically notice that you are essentially deregistering your expensive intelligence. And knowing this, the uncomfortable question arises: is this secret filter really for security or is it competition protection in a security costume? A company protects its business model and calls it responsibility. I have a different opinion on that. But we've already heard, well Elon Musk obviously did that too, it's now even become public, that he trained Grog with the help of Opus. So Opus is basically the father of Grog, of the Mecha-Hitler. So, I have to take a quick sip of coffee. And so to the timing, because this is perhaps one of the most revealing parts of the story. On July 4th and 5th, last week, Antropic publicly warned that AI was improving itself so quickly and called for a brake pedal for the entire industry. What exactly the warning is about, yes, I'll save it for more detail in the deep dive, because that will then be my big topic. But just to record the timing, on June 5th the company publicly called for a brake and on June 9th four days later the same company brought the strongest AI model in its history onto the market, bookable via any credit card, anyone can do it. So TechCrunch relished putting this in the headlines, essentially. Anthropic is releasing its most powerful model days after warning that AI is becoming too dangerous. And why this timing? A look at the calendar is enough. Antropic confidentially submitted its application for access to the stock exchange on June 1st, i.e. the S1 filing. This is this registration document for the US Securities and Exchange Commission (SEC). The valuation that is in the room, at least that was the last valuation, is 965 billion, or just under a trillion. I think at IPO it will probably be more in the same segment as SpaceX, so 1.5 to 1.8 trillion. And the IPO is likely to come in October. And suddenly it all forms a picture. So the release is also a message to investors. Look, we have the best model in the world. And double the price is the second message. Look, we can actually make money with this too. Yes, it's kind of like a demonstration of power. And so, before we get to the Apple keynote, the part that is perhaps the most important thing for you in the company is data protection. In the terms of use of the new Mythos class there is a sentence with, yes, a small problem. So all data traffic with Fable 5 is mandatory for 30 days for security monitoring purposes. And now it's coming, also for customers who have contractually agreed Zero Retention with Anthropic. So the promise that nothing will be saved. This promise simply does not apply to the Mythos class. According to Anthropic, training is not done with this data. It's fair to say that. But if you are a company that has assured its customers in order processing contracts, i.e. the contract according to the GDPR, which regulates how a service provider handles personal data, that nothing will be saved when using AI, then you are not allowed to use Fable 5 as of today. So check this before incorporating the model into your processes. This is the most concrete homework for everyone who wants to use it for this week. And at the very end, two sober footnotes so that this doesn't become an infomercial. First, the model is not flawless. Heise reports that although it wins 11 out of 13 in the comparison test in some tests, it hallucinates more even than its predecessor. So the strongest model is not automatically always the most reliable. And secondly, Meter, this is this independent testing organization that checks new models for dangerous autonomous capabilities. This also occurs again in the deep dive. They say that this model cannot yet reliably carry out research projects lasting several weeks completely independently. So that's still a big step, there's still a big step away, to this science fiction where you just tell the AI ​​and it takes care of it. And my practical rule of thumb for you, for long, complex independent tasks, i.e. large-scale analyses, entire projects, double the price can actually be worth it. But for everyday stuff, Opus and Sonnet are still the best economic choice. And if you use your passion via the API, you definitely have to deal with prompt caching or something like that. This is essentially a kind of temporary storage mechanism that makes repeated entries 90% cheaper. At these prices, it's not a free choice, but something you should bring in as a duty. So make sure you have your agent harness clean and prompt caching installed. Yes, that's the story of the week, the best model in the world that you can only get filtered, that is removed from the flat rate after two weeks and that was released four days after the warning about exactly such AI. And now you maybe understand why the story actually comes from my biggest... So, yes, you understand why it's all so strange. So, let's come to WWDC, in a nutshell. I had originally prepared a really big Apple blog for today, but Apple held its biggest keynote at WWDC on June 8th. It stands for World Wide Developer Conference, the annual developer conference at which Apple shows what's coming next year on the software side. And this year the completely new Siri. Oh, miracle! Yes, let's be honest, the longer I looked at this keynote, the clearer it became that it was a total nonstarter. Lots of shows, lots of demos, but nothing, almost nothing, that we hadn't already told here. That's why I'm going to ask you to summarize this now. Three points. The week in Cupertino didn't really provide any more real substance. First, yes, the new Siri runs on a Google model at its core. A Gemini built especially for Apple with around 1.2 trillion parameters. And Apple is reportedly paying Google about $1 billion for it. But that's exactly what we announced weeks ago, when it was still a rumor. Now it's officially nice, but not a novelty, but a confirmation. Pick it up. By the way, if you read 1.2 trillion parameters, you could very quickly assume that this is probably Gemini 3.5 Flash, because that is exactly the number of parameters that was mentioned with 3.5 Flash. This means that the focus is more on speed and not on know-how. Point 2. In the future, you will be able to choose which AI your Siri uses in the background. So Cloud from Antropic or JetGPT from OpenAI, Gemini or Glock from XAI. Apple calls these extensions. But if you think about it for a moment, the sound of evolution was completely foreseeable for two reasons. Firstly, regulation, requirements such as the Digital Markets Act in the EU. So that's the law on digital markets. This forces large platform companies, the so-called gatekeepers, to keep their systems open to competitors. Apple has to keep it open anyway, so to speak. It's better to turn an obligation into a marketing feature. And secondly, Apple doesn't actually lose anything. On the contrary, all AI providers run through Apple's App Store world. And Apple earns money from every transaction in the App Store. So Apple opens the door and at the same time collects money at the door. And point 3 is the only point that really made me pay attention because it confirms an old thesis of ours. Namely, Apple is becoming a private cloud provider. Briefly to explain, Apple operates a so-called Private Cloud Compute. This is Apple's isolated, locked data center, which supposedly not even Apple itself can look into. And the new Siri is built so that the Google model also runs within this locked space. So Google supplies the brain, but doesn't see your data. Now think this through. I predicted the same thing here in the podcast last year. And I'm patting myself on the back now. This is Apple's actual AI business model. Today Apple still pays Google a billion to get the good model. But this is just the transition phase. At the end the cash flow turns around. Then the AI ​​laboratories, i.e. Google, pay OpenAI, Tropic, and Apple so that they can access the hundreds of millions or actually two billion iPhone users via this secure private cloud. Apple doesn't have to build the best model. So Apple essentially owns the door to the customer and becomes a murder station. And this exact pattern is now officially visible. So that means they simply build a private cloud and thus become a hyperscaler, but a premium hyperscaler. And their customers are the AI ​​laboratories and not just any end customers. The rest of the keynote, i.e. faster apps, AI image editing via voice, a dedicated Siri app, yes, is all blah blah. Another side note. Yes, by the way, that was Tim Cook's last big Apple keynote. So Tim Cook is the former Apple boss. John Turnus will take over on September 1st. Yes, it's basically an end of the era. And please take one thought from the Apple topic with you, also in this short version. AI is moving into the end device. So it will soon no longer be an app that you open, but rather it will be located in the operating system. Always there with access to emails, calendars, contacts and photos. So anyone who has Apple devices in their company should ask themselves before the fall who actually controls which request leaves the device and can I as a company specify which AI my employees can choose. So if your employees use Glock and your company data and emails are then sent to Glock, that may not be what you want to do as an employer. So now we come to hard economics. That is the next topic. As I said, it's about money. It's about what AI actually costs. And the answer is clear, significantly more so this week. And above all, much more incalculable than most people think. You heard it earlier with Fable 5. The best model is out of the flat rate. And this is exactly the pattern that runs through the entire industry. So let's start with something that sounds like boring small print, but is actually a small earthquake. Microsoft has announced that it will completely change the billing for GitHub Copilot. For those who don't know, GitHub Copilot is one of the most used AI tools for programmers worldwide. This is, so to speak, the AI ​​assistant that helps developers code, suggests the code, completes it and so on. So it's definitely quite common. And so far it has worked like a Netflix subscription. You pay a fixed amount per month and then use it as much as you want. This is called a flat rate or, in technical terms, a software-as-a-service flat rate. So SaaS, software as a service. So you can use the software, so you don't buy the software, you rent it monthly. And the nice thing about a flat rate is that you know exactly what it costs. Fixed price, done. And that's exactly what's over now. So from June 2026, Microsoft will switch to a usage-based model. This means you pay for what you actually use. And per token. So let's explain again. A token is basically a word building block. AI models don’t read and write in words, but in smaller chunks. So a word like hardware maintenance is three or four tokens. And every time the AI ​​reads or writes something, it costs a token. And tokens cost money. That is, what is happening right now? The predictability of IT budgets simply disappears overnight. Before you knew I pay x per month. Now you know I pay, yes it depends on how much my people use now. And now comes the nasty part. Of course, companies are now urging their employees to use AI everywhere and become more productive. And yes, use AI, use AI, use AI. There's even a term in the industry, so to speak, called token maxing. So maximizing token usage. And the result is that costs are literally exporting. There are actually already companies that are considering layoffs because of this. Not because AI replaces people, but because the AI ​​bill is so high that they have to save elsewhere to pay it. And that's not absurd, is it? You save on staff to pay your AI bills, which you then purchase to save on staff. It's a bit like the cat biting its own tail. The technical term, which is important by the way, is called inference costs. So inference is when the AI ​​actually works, i.e. when an answer is generated. Training, i.e. building the model, is one cost side, but inference, i.e. ongoing use, is the other side and is incurred with every request. And it is precisely these inference costs that are the big unsolved problem right now. I have to make this more concrete with an example so that you can get a feel for it. You have a small engineering office, 20 people, and they all use an AI coding assistant or AI writing assistant. With the old flat rate you would have said, okay, 20 licenses, each 30 euros per month, makes 600 euros. Complete. The budget now says nothing can happen. And things are different with a token model. If one of your people comes up with the idea of ​​having AI adapt a huge document or analyze an entire project, then a single such request may burn as many troops as hundreds of small ones would otherwise. And you're standing there, yeah, and you're not even seeing it in real time. You see this on the bill at the end of the month and that is exactly the problem. A predictable fixed cost item becomes a variable and a fluctuating item that you can hardly estimate in advance. It's like going back from a cell phone flat rate where you paid per minute or SMS. Yes, just that the costs can be significantly exported. So it's basically like turning on roaming in South Africa. Yes, what are companies doing about it now? They build the so-called AI gateways. So a gateway is a gate, a checkpoint. So you could do that. There is already software for this. This means that they put a so-called control layer between their employees and the AI. She looks, is that a simple request? Then I send it to a cheap model. Is that something difficult? Only then can the newer model be used. And they set hard budgets for each department. So this is basically a type of cost control that is familiar from every other area. Except now it's made for AI queries. By the way, this has actually become a business model. So these gateways now exist. And of course you have to pay for them too. And with all this cost pressure, now things are getting interesting again for us in Europe, the cheap providers are suddenly rising. So above all Diepsig. This is a Chinese AI company. And Diepsig was the fastest-growing software provider ever in June. Because US companies are looking for cheap alternatives. And as a practitioner, I have to briefly raise my finger and say that US companies send their data to a cheap Chinese model. That may be okay in the US. But for us in German medium-sized businesses or in Austria with GDPR data protection, this is a bit tricky. So if someone says to me, Hey, Diepsig is a lot cheaper. Yes, that's right. But think carefully about what data you send in, because it may end up on servers over which you have no control. And honestly, that's just not an option for many applications. No matter how cheap. But there are also legitimate efficiency gains. For example, there is a tool called Cursor Composer. We've already talked about that. Cursor is supposed to be bought by SpaceX. And in independent tests it is almost as good as the more expensive models in terms of quality, but only costs a fraction. And that's the real message behind it all. It's no longer just the smartest model that makes the difference on the market. The ratio between smartness and price decides. Pure intelligence is becoming a mass product, i.e. a commodity. And that's a huge shift from a year ago, where everyone was just looking at who had the luckiest model. Today the question is, who has the luckiest model for the cheapest price per request? So, that brings me to the second part of the money block. And this is a story that made me look again while reading it. Namely Google. So before I go there again. Fable 5 has now set completely new standards and has set completely new enablements. To us, it may feel like you have to use Fable 5 now, no matter what it costs, because it can do so much. That's correct now. But in three months there will be models, open source models, that will have the same capabilities and will be significantly cheaper or maybe even run locally. So yes, the most important, the largest model is relevant, is also important, but the market is now different, especially if Anthropic starts now, so to speak, converts these costs into API costs and this incalculability, then that will change quickly and as soon as the first open source model is out that has similar capabilities, that will change immediately. So don't necessarily jump on the hype straight away, take advantage of it while you can. So if you have a flat rate, let it glow. Fable 5, throw all the big projects at it and see that it solves everything it can. But don't depend on it and wait, there will be an alternative in three months, at the latest in six months and maybe build up an Agent Harnest with local models, because you don't have to give everything to the smartest model. So now we really come to the second block and that is Google is now buying computing power from Elon Musk and SpaceX. So they are now renting space from SpaceX. It sounds strange, and it is, but I think I can explain it clearly. Google is one of the richest technology companies in the world. This year, Google alone invested over $180 billion in expanding its own data centers. That is a number that is hard to imagine. And on top of that, Google has announced an 80 billion share sale to build even more. By the way, they took in 85 billion. And yet, still it's not enough. Google can't keep up with the farmer because the demand for AI is simply exploding. And why can't even Google keep up? That is the actually interesting point because it is, yes, yes, because it is fundamental, because it shows something fundamental. It's not the money, it's the physics. If you order the latest Nvidia graphics cards today, i.e. if you order them today, they won't deliver them for another year. And that's not even the smallest problem. The big problem is electricity. If you want to connect a new, huge data center to the American network, for example, that alone, i.e. the network connection, will take between five and ten years. Five to ten years just to get electricity. And of course Google can't wait that long. So Google prefers to rent instead of building it itself. And now the numbers on the deals. Google will pay SpaceX $920 million every month from October 2026 to June 2029. Every month. In return, Google gets access to around 110,000 Nvidia GPUs in the Colossus data center in Memphis. In total, this deal has a volume of approximately 30 billion. And that's not even the biggest one. As we already know, Antropic pays SpaceX 1.2 billion a month for data center capacity in the same cluster. That is, what is happening here? SpaceX, the rocket company, has overnight become one of the most powerful AI data center landlords in the world. So the XAI takeover wasn't such a stupid idea. And now we're making over $2 billion a month from this compute rental. Annualized, that’s over 26 billion. For comparison, SpaceX as a whole group only made 18.7 billion in sales in 2025. This means that this new rental business is already larger than the entire previous group. Crazy, right? And of course the whole thing prepares the ground for the SpaceX IPO, which is imminent. The pricing is done today and it starts tomorrow. But the valuation that's out there is still $1.75 to $1.8 trillion. Billionmen, we've already talked about it. And when the deal was announced, it immediately sent shockwaves. So the shares of other GPU rental companies like CoreWeave have fallen by 7%. Even Nvidia is down over 6% as investors get nervous about the concentration of power at SpaceX. Now you have to tell something familiar that is completely overlooked in most reports. But once you know it, you see the whole deal differently. Alphabet, Google's parent company, itself owns a stake in SpaceX. So around 6 to 7% are reported. This goes back to an investment in 2015. So Google doesn't just rent computing power from another company. Google is both a customer and co-owner of SpaceX. And that means that if Google now transfers $920 million per month to SpaceX, then Google is inflating the sales of a company that owns a stake in Google itself. And that just a few weeks before the IPO. I think everyone understands what's happening. Google has a vested interest in making SpaceX look as plump and healthy as possible for its IPO. And the higher the SpaceX valuation, the more Google's share is worth. This isn't illegal or anything. I don't even want to say that. But you just have to know that because it puts the deal in a different light, of course. Part of the turnover that justifies this huge valuation is essentially generated by a shareholder himself. And that's a bit like buying your own house from Kinko shortly before it was sold by your brother and it's going great in the exposé, fully rented out. Yes. After the announcement, I of course calculated SpaceX's valuation again. We already did that two episodes ago. Is this rating of 1.75 actually justified? And if you measure it against what comparable companies are worth. And now I have to briefly explain in which league SpaceX is now playing. Namely the league of the so-called neo-clouds. So new cloud providers that basically do nothing other than rent out AI computing power. The best known of these is called CoreWeave and is also on the stock exchange. And then there are the old giants, the hyperscalers. These are Amazon with AWS, Microsoft with Azure and Google Cloud. So the established large cloud companies. But now the numbers. And they are, yes, illuminating, you could say. The entire neo-cloud market, i.e. all these AI rental companies combined, had sales of around $25 billion in 2025. The entire market. And now guess how much SpaceX makes from its AI rental company alone. Over 26 billion. This means that SpaceX single-handedly makes more sales with AI infrastructure than the rest of the neo-cloud market. That's quite a number, you have to acknowledge that. So for comparison, the best-known landlord, according to CoreWeave, makes around 5 billion in sales. So SpaceX makes five times that. And even Oracle's infrastructure is only 18 billion. So the sales are real and huge. And there is simply nothing to shake about in that. But, and now we come to the exciting part, now it's about the evaluation. The question is not how much sales, but rather how much the company is worth in relation to this sales. So there is a key figure for this, which is called Price to Sales Ratio in English. This simply says how many dollars of market value do I pay per one dollar of annual sales. And that's where it gets interesting, CoreWeave, the direct comparison, is worth about 11 times its sales on the stock market. So price sales ratio of 11. The large established tech companies like Amazon, Microsoft, Google are only between 3 and 8. And now SpaceX, 1.75 trillion valuation, divided by 26 billion sales. That gives a price turnover ratio of 67. That means if I'm generous and include all of SpaceX's turnover, including Starlink and the rockets, I still get around 40. This means that SpaceX will be valued at around six times as expensive as CoreWeave when it goes public and a good five times as expensive as a normal hyperscaler, measured in terms of sales. And what does that mean in plain language? If you were to value SpaceX only through the sober lens of an AI landlord, i.e. with CoreWeave standards, then the business would perhaps be worth 200 to 300 billion, not 1.7 trillion. That means the other trillions that are in this valuation are pure future fantasy. This is the bet on Starlink, on Mars, on Starship, on the idea that SpaceX will one day have a monopoly on space. And my sober conclusion, as someone who likes to look at numbers, AI sales are great for the shop window. It rounds out the story, it delivers the best, i.e. the nice recurring income. Investors love this, but it doesn't justify the moon valuation. This review is not a review of a GPU rental, this is purely a matter of faith. And everyone has to decide for themselves whether they believe in this belief. I only say 67 times sales. This is very, very sporty. That doesn't mean that it won't happen again at some point. You could now add the purchase of cursors, which adds another 4 billion. If you add in the growth of SpaceX or the other business, I think that's another 30 percent of 18 billion. So here we come again with 5 billion. So then the multiple comes down to maybe 30. But it's still a very sporting affair. But what you have to say is that it is extremely capital efficient because XAI has given out, I think, 45 billion and they can now get the coal back in just under 2, so 18 months or so. And if they now had enough capital and expanded data centers at this speed, then it could actually become an interesting business. But there are also many other players. So another thing of megalomania. Elon Musk presented plans for orbital data centers at an investor pitch moderated by, of all people, Jamie Dimon, the head of JP Morgan, one of the largest banks in the world. So orbital, i.e. in space. Computing clusters on satellites that run on unlimited solar energy and thus simply circumvent the entire electricity problem on Earth. Because I don't know whether to find this brilliant or completely crazy. Probably both. But it just shows how serious this electricity problem actually is. And the lesson from the whole blog with this money, very practical for you. Electricity and computing power are essentially the new oil. Whoever has the secured power connection and the chips has the power. And for anyone who relies on the cloud to be available, don't blindly rely on the fact that there is unlimited capacity. So that's not it. This is actually something that should not be forgotten. So now we come to my favorite section, KI's Next Top Model, where we look at a new one, i.e. where we look at the new models. And yes, we already had the giant model of the week with Fable 5 in detail in the first blog. But there were also a few things that came out at the completely other end of the size scale that I find almost as exciting because they show a counter-trend. And the trend is that small, smart and efficient beats huge and expensive. To be honest, this is exactly the opposite movement to what I just told you in the money blog and to the Fable price list anyway. Yes, we need to take a quick sip of water. Let's start with a model that I find really impressive. It comes from a startup called Syfra out of San Francisco. So the model is called Zaya 1 8B. So that means 8 trillion, so 8 billion parameters. And what’s special is the architecture. This is a so-called Mix of Experts, or MOE for short. In German this means a mixture of experts. And I'll explain this with a picture. Imagine that you don't have one all-rounder who has to answer every question, but rather you have a team of many specialists. And if a question comes up that only one person needs, i.e. the right specialist, only the right specialist will be woken up and he will then know what to do and the rest will continue to sleep and cost no energy. The Mix of Experts works in exactly the same way. The model has a total of 8 billion parameters, but when it actually works, it only uses around 760 million of them. So not even a billion active. Yes, and why is this such a big deal? Because the model is so small and economical that it runs locally on inexpensive devices without any problems. Even on a smartphone. There is no need for a huge data center. And yet, hold on to your hats, this tiny model beats many, many, much, much larger models in independent tests. For example, it proposes a model called Mistral Small with 119 billion parameters. So more than 15 times more. So the other model has 15 times as many parameters and still loses. And on a difficult math test called HMMT, Zaya 1 achieved 89.6 percent, even beating Klozonet 4.5, which scored 88.3. And it also produces good peak values ​​when programming and when logically closing. By the way, this is called intelligence density. So how much cleverness can I manage per parameter, per calculation unit? And the lesson is, the dull one, a model has more parameters than yours then loses meaning. It's no longer about who has the biggest engine, but who builds it the cleverest. And to be honest, I think that's a healthy development because it means that you can achieve really good AI even without millions of data centers. And think about what that means in practice, especially for us here in Europe, especially for medium-sized businesses. If a model like this is small enough to run locally on a normal computer or even a cell phone, then that means your data doesn't have to leave the device at all. You no longer have to go to any American or Chinese cloud, you don't have to send anything to a cloud. And that, if you ask me, is the actually good news in all this AI madness. While the big companies are outdoing each other with trillions of data centers, the smaller, efficient models are giving data protection-conscious companies the opportunity to use AI without giving up control of their data. That doesn't solve all the problems, of course, but it's a direction that I think is much healthier than sending everything to the largest provider and hoping that it goes well. And that brings me to the next one, which fits thematically perfectly. A company called News Research came out with something called Hermes Desktop. I've talked about that before too. This is an AI agent that you run locally on your own computer. Can be used on Mac, Windows, Linux. So agent memory is an AI that not only responds, but also completes tasks independently. And the exciting thing is that the underlying agent is open source. Anyone can look in and use it. It only launched in February 25 and has already collected over 180,000 stars on the developer platform GitHub. It's kind of like liking developers. And therefore the fastest growing agent framework in the world. But now comes what I find technically really clever. Most AI agents are, as they say, stateless. This means that every task starts from scratch. The agent doesn't remember anything. No matter whether he has already solved this problem 100 times, the 101st time he starts again from the beginning. It's like waking up every morning and having to learn how to make coffee. Hermes does exactly that differently. Hermes has a so-called closed learning loop. After each task the agent checks, did it work? How did I solve this? And then he saves the solution as reusable abilities, i.e. skills. Quite simply as a text file in a small local database. And when a similar task comes up later, he digs out this saved skill instead of calculating everything anew and expensively, i.e. recalculating it. And the result is measurable. Agents who have developed more than 20 such skills complete new tasks on average 40 percent faster and more economically. That is, this agent gets better over time because it learns from its own experience. And the whole thing runs with a really graphical interface, with security barriers, so people without programming knowledge can also use it. I think this is a bit of a foretaste of where we're headed. Away from AI in the cloud towards AI that runs locally and learns with you. Speaking of AI improving itself, there's a third little puzzle piece that points in exactly that direction. And I'm deliberately saving that here because it will become important in more detail later. So a Japanese startup called Sakana AI founded its own laboratory in June, the Sakana AI RSI Lab. And at this point I have to briefly interject, because Sakana is an old acquaintance for us, who was on the table for the first time last year, who was on the table for the first time over a year ago, so in May, June last year, then again in November and also this spring. This topic has been going on with us for quite a while and that's exactly why I find it remarkable that they are now opening their own laboratory just for this purpose. They're pretty serious about it. So RSI Lab, that stands for Recursive Self-Improvement. This is the technical term for AI, which continues to improve itself. So remember the word, it will come again in a moment. And the exciting thing about Sakana is the approach. So they don't just want to improve AI through ever larger data centers, but through evolution, i.e. through a principle that they have copied from nature. You create many variants, let them compete against each other, the best survive and are further developed again and again. This is basically Darwin for AI models. And now, in retrospect, the name Darwin-Gödel machine that we talked about back then also makes sense. Instead of using raw computing power, you use selection. And I think that's worth mentioning because it shows that there are people who are consciously looking for a way out of this ever bigger, ever more expensive arms race that I already told you about on the money blog. And before we go any further, just one more thing that, although not a new model, is strategically huge. So OpenAI, the company Interjet GPT, has announced and now I quote, chat is dead. So chat is dead. That's what a high-ranking AI man told the Financial Times, meaning that the era in which you type a question and get an answer is coming to an end. Instead, OpenAI is completely rebuilding Chat GPT. The biggest renovation since the start in 2022. The chatbot is to become a so-called super app. So a super app is an app that bundles a lot of services under one roof. So the model is WeChat from China, where you chat, pay, shop, order a taxi, everything in a single app. Just like the general purpose app, Chat is intended to be GPT. With AI agents, with programming tools, with image generation and with connections to companies like Canva or Booking.com. Let's just take a quick sip of coffee. The numbers behind it are quite impressive. So OpenAI programming tool Codex has increased sixfold since February to over five million weekly users. And on June 3rd they opened it not only to developers, but also to product managers, lawyers, analysts, even people who don't program at all. Two million companies now use OpenAI tools and they account for 40 percent of sales. There is even a new lockdown mode that restricts companies' web browsing and agent mode so that nothing uncontrolled happens. And that is exactly the point that interests me the most. So if AI agents act independently in your system, in your CM, in your accounting, then you need guardrails. Then you need clear rules about what the AI ​​is allowed to do and what it is not allowed to do. So the era of chatbots for employees is coming to an end and the era of AI is beginning to carry out multi-stage tasks independently and that is changing quite fundamentally how you introduce AI in a company. And you know, and you know what, so let me tie the whole model block together with a clear appeal because I think that is the most important practical message today, especially for us in Europe. Look at what's coming together here. So on the one hand, we've seen the costs of AI from the cloud explode. The top models make your bills incalculable and there are even companies that are considering layoffs because of this. On the other hand, we can see right here in this blog that there are now smaller, more highly efficient models like Zarya 1, which runs locally on your own computer and there are open tools with Hermes Agent with which you can use this even without programming knowledge. And for me the conclusion is crystal clear. Now is the time for every European company to get serious about local inference solutions. So inference, you remember, this is the ongoing use of AI. And local simply means on your own hardware in your own house instead of in some American Chinese cloud. Start experimenting. Take a small open source model, take an open harness, i.e. an open source control framework like the one from Hermes that we just had, or something like OpenSpec and try out what works with it in the company. Why? Quite simply, so that one day you don't have to make the absurd decision that I described earlier, namely having to lay off employees just to be able to pay your AI bills. If the AI ​​runs locally at your location, then the bill can be calculated. Your data stays in-house, you are not dependent on an American provider that will triple its prices or change its billing models tomorrow. This is honestly one of the most sensible strategic moves a medium-sized company can make this year. Don't outsource everything, don't blindly follow the hype, but build up expertise in-house, start small, try it out now. Not just when the first shock bill comes. So now comes the part that I'm looking forward to, not looking forward to maybe isn't the right word, but a part that's been on my mind the most this week, my deep dive. The company Anthropic, just like Anthropic again, Fable 5 company from Block 1, published a report on June 4th entitled When AI Builds It Self in German. When AI builds itself, mind you, five days before the Fable release. Yes. And this isn't a marketing paper; on the contrary, it's a pretty honest, almost reassuring self-disclosure. Anthropic has disclosed internal numbers from its own server rooms and shows this recursive self-improvement, you remember Sakana's word, RSI. AI that improves itself is no longer a dream of the future. It has already started. And now the numbers that really got me thinking: in May 2026, over 80% of the entire program code that was built into the real productive systems at Anthropic was no longer written by people, but by AI Cloud itself. 80%. And if you count experimental things, if you include all of these experimental things, it's even over 90%. For comparison, before Anthropic launched its Cloud Code tool in February 2025, this figure was in the low single digits. So before it was a few percent, now over 80. In just over a year. This means that the human engineers at Anthropic hardly write any code themselves anymore. They are basically just directors who tell the AI ​​what to do and then control the output. One of the top engineers said in an internal survey that he hasn't typed a single line of code himself in five months because cloud simply makes the work error-free. And productivity has exploded. The average Anthropic engineer now delivers eight times as much code as they did just a few years ago. And in a survey of 130 researchers, they said their scientific output quadrupled. But now come the numbers that really back it up. There was a test in which the model optimized its own training code to make it faster. A year ago in May 25, Cloud achieved an acceleration by a factor of 3 in the test. The model that was still running internally under the code name Mythos at the time, i.e. the exact model that we have been able to book for today in the filtered version as Fable 5 since yesterday, achieved an acceleration by a factor of 52 in exactly the same test in April 26. So a factor of 52 in relation to a factor of 4. A highly qualified person, i.e. a human researcher, would need around four to eight hours of work to accelerate to a factor of 4. The AI ​​does many times this in a fraction of a second. And it continues. The success rate for really difficult open programming tasks rose from 26 percent to 76 percent within six months. And the amount of time that such an AI can work independently without human intervention is still growing rapidly. So in March the tasks lasted around four minutes. Today the model can process tasks that take twelve or 16 hours. An independent testing organization, i.e. the tenants, confirmed that Mythos can work autonomously for at least 16 hours without errors. Error free. There was even an experiment where two experienced human researchers solved 23 percent of the research problem in an entire work week. The AI ​​agents solved 97 percent of the same task. They independently formulated new hypotheses, programmed a test environment, carried out and evaluated the experiments. The whole thing cost about $18,000 in computing costs. And to be honest, $18,000 for a machine to solve 97 percent of a research problem that two professionals could only do a quarter of in a week is, economically speaking, a bargain. And that's exactly what makes it so easy for Ork. But now comes the reason why the report is not a joyous announcement. So a trope warns about itself in this report. They say that the bottleneck, the real bottleneck in AI development is no longer writing code, the machine does it. The bottleneck is the human reviewer, i.e. human checking of code. Because if the AI ​​soon writes code better than any human and then Tropic says that will definitely happen next year, then we humans will no longer be able to control this code quickly enough or deeply enough, cognitively or temporally. All we can do then is watch and hope that it happens. And I want to make this tangible with an analogy like this because it's such an abstract point, but incredibly important. So imagine you are an editor at a publishing company. You used to read and proofread one book a year, calmly, word for word, and now your author, the AI, is suddenly writing thousands of books a year and you have to check and approve them all. What happens? You can't do it. You don't read everything anymore. You start to fly over, you wave through and at some point you reveal something that you actually no longer really understood. Simply because it's too much and too fast. And that is exactly the danger with the AI ​​code. It's not that the machine is becoming malicious, but rather that we humans are losing track of things because we can no longer keep up with the pace. And the thing goes into production and no one has checked it and no one knows what is really happening. In fact, in parallel, Antropic publicly called on June 5th for the industry to develop a brake pedal, but a brake pedal, that is, a technical mechanism with which to improve the introduction in front of itself, which can slow down or stop AI. Co-founder Jack Clark said that around 80% of the programming work at Antropic is already done by itself. And in a few years the direction could be 100%. And they even argue for the possibility of a coordinated global development pause. Yes, with the honest admission that a break from a single laboratory would be quite pointless because the others will continue. And I thought that was a powerful image. They say it's much harder to control than nuclear weapons because you can't see an AI from satellites the way you see missile silos. AI training simply runs invisibly in some data center. Now put that... Yes. Yes. Yes. And today? Yes. Today they are launching the whole thing. So that's a bit of a hypocrisy. That's right... And if you put that aside, why are they doing that? I'm going to play devil's advocate now. So the opposite position, because I think you shouldn't just swallow the whole anthropic, the good, responsible, Mahana narrative. So we've already discussed that several times. Just ask yourself the simple question: who actually benefits from a break like this? And then look at who's calling after the break. And the company that is currently at the forefront with the best or one of the best models and is currently preparing its own IPO in October is calling. Now think about it, if everyone else actually took a break, who would benefit the most from it? Well, exactly. The one who is already at the front. Anthropic, so to speak, freezes the race while they are in the lead. Cement break every day, so they cement their position as number one. The competition is not allowed to catch up and Anthropik sits comfortably at the top and can go through with its IPO in peace and quiet. Yes, that would be like shouting in the middle of a race to a runner who is currently in the lead, guys, let's all stop for a minute and talk about the dangers of running. Yes, of course, comfortable when you're in front. I'm not saying the worries are made up. The numbers are real and impressive anyway. You also have to read the message through the lens of self-interest. A break that just happens to benefit the person asking for it is always a bit suspicious. And I think that's part of it if you want to be honest. Not everything that comes under the banner of security and responsibility is free from business calculations, I'll say it now. There is another human side that I want to mention because it struck a chord with me. The report says there is something of an existential crisis among engineers. So the whole culture of colleagues helping each other, hey, can you take a quick look at the code, my code. You sit together, debug something at the next table, yours, a kind of small gift economy between colleagues. This disappears completely because the machine does everything immediately and sterilely. And some employees report a feeling of insignificance because they lose touch with the systems they are supposed to monitor. And I honestly think that’s one of the most underestimated aspects of the whole AI story. We always talk about jobs that are being eliminated, but we don't talk enough about the meaning that is being eliminated. Even with the people who keep their jobs. But now it's getting exciting because here comes the contradiction that hasn't left me in peace this week. On the one hand, we have this company Antropic, which warns, which calls for brake pedals, which calls for regulation, the good, careful, responsible AI company. And on the other hand, in the same week, reports come out that this same Antropic has embedded some engineers directly into the secret division of the NSA. NSA is the National Security Agency, the largest secret service in the USA, which is responsible for interception and cyber operations. And there's even a technical term for this sort of thing, which I don't want to reserve today because it's so revealing. These people are called forward-deployed engineers. That's actually military speech. Forward-deployed is the term used to describe troops that are sent directly to the front. And that’s exactly what Antropic does. They send their best minds straight to the cyber front, right into the secret service. And what are they doing there? They take this myth, i.e. the unfiltered full version, of which we mere mortals have only been able to get the Faible Five version with the security filters since yesterday. The same model that I just described to you with a factor of 52. And use it as an offensive cyber weapon against the critical infrastructure networks of geopolitical rivals, especially China and Iran. In concrete terms, this means that the AI ​​searches for security gaps, so-called zero days, in the opponent's IT systems in real time. So Zero Day is a language that no one knows yet and for which there is no protection yet. Hence the name. The defenders have zero days to close it. And the AI ​​then fully automatically generates the appropriate attack code, the so-called exploit, to get through the firewall, i.e. the enemy's protective wall. And understand what that means. So the same superhuman abilities that Cloud uses to write code and solve research problems in seconds can be repurposed to find and exploit tens of thousands of vulnerabilities in a country's digital infrastructure in seconds. Human defenders in the security centers have no chance against something like this. They are completely overwhelmed by the speed and complexity. And now the circle closes to Block 1. You remember Project Gleiswing, the program through which selected partners get access to unfiltered versions of Mythos. Officially for cyber defense and critical infrastructure. We have already spoken to Gleiswing in the podcast, several times. This destructive potential has already become apparent. An AI that penetrates deeply into networks, separates the noise from real signals and systematically undermines the protective mechanisms. This is no longer a theory. This is tested. And that's why the friendly word access program doesn't really warm my heart. And that needs to be brief at this point. I think this is a point where we shouldn't just move on to business as usual. We're not talking about slightly better antivirus software here. We are talking about the possibility of a single machine being able to breach the digital defenses of an entire country, fully automatically and in real time. And that dramatically and irreversibly shifts the balance of power between states. Whoever has this best AI cyber weapon has an advantage that is hardly comparable to classic weapons because it is invisible, because it works immediately and because it is difficult to trace. And this is so explosive that this incident is now officially listed as a serious AI incident in an international risk database, i.e. the OECD AI Incident Database. Because there is a risk that such an autonomous attack will get out of hand and hit civilian infrastructure. Electricity, water, banks, whatever. And now comes what I actually call paradox. And I have to break this down for you very clearly because it's easy to get confused. There's a second military story on Unchopic this week. It goes in exactly the opposite direction. The Pentagon, i.e. the US Department of Defense, is currently actively stealing from its secret military systems. Why? Because Antropic refuses to remove certain protective barriers. So Antropic has two red lines from the start. Firstly, no mass surveillance of US citizens and secondly, no fully autonomous weapons that target people without human approval. And Antropic is sticking to this red line, and the Minister of Defense or Attack Minister, as I believe he calls himself, Pete Hexef, has even declared Antropic a supply chain risk, we already had that, and is now looking for an alternative at OpenAI, Google or XAI. Eight other companies are approved for the secret systems. But Antropic is not one of them. And now the question arises, do you perhaps understand the paradox? On the one hand, Antropic is even suing the Pentagon for blocking Glot in harmless administrative areas. And on the other hand, Antropic's best people are in the deepest, most secret level of the NSEE, building cyber weapons. So it's too safe for the Pentagon, but cyber weapons for the NSEE doesn't really fit together. And I've thought about it for a long time and I think the honest point is, it's not necessarily a direct contradiction, even though it sounds like it. They are two different authorities. The Pentagon is military, the NSEE is intelligence. They are two different models. The normal toilet used in the Pentagon and the secret myth at the NSEE. There are two different approaches, so missions, and that is, one is autonomous weapons on the battlefield, the other is intelligence, cyber warfare. But that's my honest point, no matter how you slice it apart, in the end you're left with a bad feeling. A company that publicly warns about the dangers of its own technology and demands a brake pedal is simultaneously weaponizing this very technology. And I find this tension cannot be ironed out. This should be endured and named. And now we come full circle to what I said earlier about the brake pedal. Antropic sells itself as an ethical, principled company. We don't do autonomous weapons, we don't do mass surveillance, we're the good guys. But when the same company sends its forward deployment engineers to the NSA to build offensive cyber weapons, it pretty obviously violates the very principles it touts. And that's why my advice to you, and this actually applies to the entire industry, is not to blindly believe these glossy statements from responsible corporations. Not because they are all evil, but because behind each of these appearances there is tough business and tough geopolitics. They warn of the dangers, they demand a break, they insist on the principles and at the same time they collect billions, prepare their IPO and build weapons for the secret service. And this is all the same company in the same week. And it is precisely this simultaneity that you have to see in order to understand what is really going on here. Because I think that's actually the real picture of the year 2026. We have an AI that can accelerate itself under a factor of 52, that solves math at the top level, that writes its own code and that exact same machine, when you turn it over, becomes the most precise cyber weapon that has ever existed. Both are the same technology, both happen at the same time. And the people who should know best are building things up themselves, calling for a break and pushing on the gas at the same time. I think that is the most honest and at the same time most uncomfortable sentence that I have given today. So, I'm going to have another sip of coffee. After this heavy fare, let's take a step back and look at the big picture, because something fascinating is happening right now and that is, so to speak, the political answer to everything I've said today. The question that arises now is, if these AI companies become so profoundly valuable and so profoundly powerful, who should actually own the AI? And the exciting thing is that this question comes from both political camps, from the left and the right at the same time. And it's really rare that the two of them agree on anything. Let's start on the left with Bernie Sanders. You may know him, he is the senator, the left-wing man in American politics. He has introduced a bill called the American AI Sovereign Wealth Fund Act. So Soverän Wealth Fund. I have to explain that. This is called a sovereign wealth fund, which is a state investment fund. The best known example is in Norway. They have used their oil revenues to build a huge sovereign wealth fund over decades, which today holds shares in companies all over the world and receives the profits, so the entire country benefits. And that's exactly what Sanders wants for AI, a one-time tax of 50% payable in money, not in money but in stocks, on the leading AI companies, i.e. OpenAI, Antropic XAI and so on. That means the state would get half of these companies, the voting rights with one seat on the board of directors for each company. And the stated goal is that the state should be able to block decisions that harm citizens and that the profits and dividends should be distributed to the public. Sanders' reasoning is that these companies have been trained with publicly available data, with our texts, with our images, with our knowledge from the Internet and thus create gigantic value. So the common people, i.e. the general public, should also get something out of it. And if you're honest, that's not such a far-fetched argument, because we learned these models, which are now worth trillions, by basically having the entire Internet at their disposal. Everything, i.e. everything that millions of people have written on the Internet over the decades. Every blog post, every Wikipedia, every forum was basically the source of raw materials. And now they are making, a few companies are making unimaginable sums of money from it. One can ask the question: is it fair that the value created by all of us from this knowledge ultimately only ends up with a handful of shareholders in Silicon Valley? I'm not saying that Sanders' 50% solution is the right one. I think there is a better solution, which is taxes. But I think the basis behind it is absolutely justified. So now the surprising part. You would think that the Trump camp, i.e. the right-wingers, would immediately shout socialism, stay away from the companies. But that's the real news. The Trump camp doesn't think the basic idea is that bad. Just on a smaller scale. On June 5th, Trump confirmed on board Air Force One that his government had been negotiating direct government participation with OpenAI for over a year. In this way, quote, the public practically becomes a partner. And a research institute, the Cato Institute, estimates that the White House already has stakes in 20 private companies. So shares, options or golden fairs, i.e. special shares with veto powers. The real argument is no longer about whether, but about the WiFi. Bernie Sanders says 50%. Sam Oldman, the head of OpenAI, says yes in principle, but 50% definitely not. The industry itself prefers something like 1-5% as a contribution to a public fund. So it will probably settle somewhere between 1 and 50%. That's a huge range, but yes. What does this mean for us now? You and I are not directly affected by this. But there is a classification that I find important. If the state becomes co-owner of an AI laboratory, this will change their incentives in the long term. So the pricing, the product strategy, the question of which countries technology can be exported to. Everything looks different when the state is at the table. This means that for anyone who strategically relies on a specific AI provider, i.e. possible partial nationalization, there is now a new risk factor that must be kept in mind. Not dramatic, but it belongs on the line. And basically this brings everything I've talked about today full circle. AI has become so powerful, so valuable and so dangerous that hands are now reaching for it from all sides. So Anthropics rations access to the best intelligence. Apple reaches out to you for the interfaces. SpaceX is grabbing the computing power. Secret services grab the clout. And now the state is also seizing property. Everyone wants a piece of the pie. And that just shows what a historic moment this is. Yes, it was funny. Well, my friends, that was a lot. I know, let me summarize this very briefly so that you can go forward with a clear head. Five things to take away from this episode. First, the story of the week Anthropics has released the strongest publicly available AI model, Fable 5. But it is the same model as the secret myth. Only with filters. It will be removed from the subscription flat rate after two weeks. And it came four days after Anthropic's own warning about just such AI. Second, WWDC was a confirmation event. So Siri runs on Google. The choice of model was foreseeable from a regulatory perspective. Apple is building a private cloud, a murder station so to speak, where AI laboratories will pay in the future. AI is finally moving into the end device. Then thirdly, AI is becoming more expensive and unpredictable. This, yes, free token usage has also come to an end. So the subsidy of tokens is over. The end of the flat rate at GitHub-Copilot is basically the harbinger. And even Google has to claim computing power from SpaceX because electricity and chips are the real bottleneck. Even if there may be other interests as well, as we've talked about. Fourth, small beats big. The most exciting models besides Faber are the small, efficient ones that run locally and remember things. And fifthly, the heavier part AI now builds its own code over 80% at Anthropic and is mentally used as a cyber weapon. And even the state now wants to own something. And a quick preview of what's next week. There's a lot coming up. This week it will be concrete on the stock market. So the IPO will be priced on the morning of the 12th, tomorrow on the 11th. And the first trading day is Friday the 12th under the symbol SPCX. And here the $1.75 trillion valuation is one of the biggest stories. That remains to be seen. Whoever wants to sign will have to decide this week. But a quick reminder, this is not investment advice. You all have to know that for yourself. We are also still waiting for two models that have been announced. then Cloud Sonnet 4.8, which probably doesn't matter now with Fable 5 and Gemini 3.5 Pro. That is still delayed. So there could be a lot more to come now. Yes, hopefully Christian will be there again next week. Then there's the usual back and forth again. And I'm curious to see what he says about all the anthropological stories, because I think we still have a lot to discuss. And now it's your turn. I came up with a pretty steep thesis today and I could be wrong. That's part of it. So feel free to tell me what you think about it. What do you see differently? What did I miss? Especially with the Fable story. Would you pay for the best model? Or is the second row enough for you? Or the question of how to deal with AI assistants in the company? I'm really interested in your practice. Write to me. I actually read all of it. That was AI and Tech to Go - the practical pitch. Solo with me today. Stay curious, stay critical and pay more attention to your data. Until next week. Take care. Bye. We'll take this. See you. Subtitling by ZDF, 2020