Austria Aims to Lure Anthropic to Europe as Chinese Models Surpass Claude Mythos

Austria Wants to Lure Anthropic to Europe

With export controls, the U.S. has effectively blocked access to Anthropic's latest models, 'Claude 3.5 Opus' and 'Claude 3.5 Sonnet,' for all foreigners, forcing Anthropic to lock out all users due to a lack of nationality verification. The U.S. Department of Commerce has since reopened access to Claude 3.5 Opus for about 100 selected U.S. companies but reserves the right to change the list at any time. Austria's State Secretary for Digitization, Alexander Pröll, proposed in an open letter to EU Commissioner Henna Virkkunen to lure Anthropic to Europe with legal certainty, market access, and capital. OpenAI is taking a similar line, initially releasing GPT-4o only to institutions designated by the U.S. government. In parallel, the EU Commission has responded with the Cloud and AI Development Act and a 'Chips Act 2.0': tripling data center capacity and doubling the semiconductor share from under 10 percent by 2030. Virkkunen warned of a foreign 'kill switch' over European services but admitted that 80 percent of the technology comes from outside Europe and that visible results are not expected until 2030 at the earliest. → en.sedaily.com

Synthszr Take: Pröll has put his finger right on the sore spot: Europe is executing foreign decisions instead of making its own. My Wahl-O-Mat has ranked Volt for years, so this isn't EU-bashing, but a sober calculation. Luring Anthropic to Vienna or Brussels with a subsidy sounds charming, but it doesn't change the harsh reality: 80 percent of the technology comes from abroad, and a foundation model provider won't emigrate for legal certainty as long as compute, capital, and talent are based in California. The leverage lies elsewhere, namely with the models over which the U.S. has no switch: Mistral, Llama on-prem, European open-source stacks. This is exactly what CADA addresses with data centers and chips, but on a timeline to 2030, which is damn long in a market that turns in months. Anyone building a production-ready agent architecture today should design it to be model-agnostic and include a sovereign layer, rather than relying on a single U.S. provider whose user list Washington could curtail tomorrow. Sovereignty is not a location subsidy, but an architectural decision, and it's one that can be made today.

Marc Andreessen: GLM-5.2 Beats Claude Mythos

Security researchers report that Zhipu AI's new model GLM-5.2 (also known as Z.ai) rivals the best U.S. models in detecting security vulnerabilities, even though it lags behind Anthropic and OpenAI in other tasks. Marc Andreessen says GLM-5.2 is the first Chinese model to catch up with and often beat the top American labs without compromise. In benchmarks from Semgrep, GLM-5.2 beat Anthropic's Claude Opus 4.8, and with additional instructions, both reach the level of Mythos. According to OpenRouter, GLM-5.2 is among the ten most-used models worldwide and, as an open-weight model, is freely downloadable and modifiable without oversight. In parallel, the Trump administration has tightened access to U.S. models: OpenAI is limiting GPT-5.6 due to security concerns, an Anthropic model was completely blocked for over two weeks, and even the NSA temporarily lost access to Mythos 5 and Fable 5. Saif Khan from the Institute for Progress calls the combination of model blocks and simultaneous chip exports to China a gift to Beijing. 360 Security's CEO Zhou Hongyi announced its own Mythos-level tool, Tulongfeng. Microsoft and others are already exploring offering Chinese models on their platforms. → www.wsj.com

Synthszr Take: The idea of containing an open model via export rule was a PowerPoint illusion from the start. Open weights run locally, without license servers, without a recall button. Anyone who downloads GLM-5.2 isn't asking for permission from anyone in Washington. In the Code Crash framework, we run DeepSeek R3 at $2/$8 per 1M tokens with the note 'compliance question,' and this is precisely where the real task lies: not forbidding, but classifying. What's fascinating is the double-edged sword that a security use case (finding bugs) is becoming the benchmark, because the same model that detects vulnerabilities can also exploit them. If the U.S. releases Claude 5 again next week while Chinese open-source models circulate freely, the cost curve for security tools will have plummeted. For anyone building architecture here: hosting GLM-5.2 or DeepSeek on Bedrock or Azure instead of self-hosting is the only viable path for medium-sized businesses, and the compliance question needs to be settled beforehand, not afterward.

Asia's AI Startups Counter Trump's Export Ban with Their Own Mythos-Class Models

Two weeks after the Trump administration blocked Anthropic's security models Mythos and the more restrictive variant Fable 5 for all non-Americans, Asian providers are filling the gap. The Chinese cybersecurity firm 360 introduced Tulongfeng on Wednesday, a tool that, according to Reuters, is supposed to rival Mythos, plus Yitianzhen for automated cyber defense. Days earlier, the Tokyo-based startup Sakana AI had already launched its model Fugu (Japanese for pufferfish), which it claims is 'on par' with Fable 5 and Mythos Preview and is built to orchestrate other models via their APIs. Sakana, founded in 2023 by Google alumni David Ha and Llion Jones, as well as former Mercari and Stability manager Ren Ito, calls the timing 'purely coincidental,' but already advertises on its website with 'Frontier performance without the risk of export controls.' Ha describes orchestration models as 'the next frontier beyond larger models' and calls collective intelligence the practical hedge against the concentration of power. 360 founder Zhou Hongyi, in turn, declared vulnerability-seeking AI a national strategic asset. Anthropic recently reported a run rate of $47 billion; how much of that depends on Asian customers is unknown.

Synthszr Take: Export controls work in a world where model weights represent a moat of maybe six months. When the U.S. government blocked Mythos and Fable in mid-June, it was the starting gun for the very providers it actually wanted to slow down. Sakana has been building Fugu since last year and presented the research at ICLR in the spring; the ban merely provided the stage. More interesting than the saber-rattling is Ha's argument: anyone who ties national infrastructure to a single provider has bought a risk that can turn into a full stop overnight. This is the logic behind 'Reasonable Sovereignty,' which should also resonate in German boardrooms, with Llama 4 and Mistral Large 3 as a production-grade backup instead of blind faith in a U.S. lab. Ren Ito is right in Evian: AI should be developed collaboratively, not hoarded. Those who bet on multi-model orchestration now are building in optionality before the next political decree forces it.

How GLM-5.2 is Triggering a New DeepSeek Moment for Agentic Coding

GLM-5.2 is hitting the market in a perfect storm: engineering teams are pulling the plug on closed models as API bills explode and return-on-investment fatigue grows. The model is a 744-billion-parameter Mixture-of-Experts under a permissive MIT license, activating only 40 billion parameters per token and delivering a functional 1-million-token context window (not a marketing promise, but actually usable). On the Artificial Analysis Intelligence Index, it leads all open-weight systems with 51 points, right behind Claude Opus 4.8 and GPT-5.5. Coinbase CEO Brian Armstrong confirmed that his team uses GLM-5.2 and Kimi-K2.7 as the default in their internal LLM routing gateway. Independent production tests by Cline and GMI Cloud show that GLM-5.2 regularly outperforms Claude Opus 4.8 in targeted bug-fixing and refactoring, with fewer logical errors. The cost calculation is drastic: a 45-minute autonomous session processed 6 million tokens for just $3.36. → AlphaSignal

Synthszr Take: $3.36 for 6 million tokens in an autonomous bug-fix session—that's the number that matters this week. In Code Crash, I described how Opus 4.7 reduced a six-figure codebase by 40 percent overnight and simply deployed it. Exactly the same delivery capability is now available under an MIT license, runnable locally on a Mac Studio with 256 GB. This shifts the compute discipline from the question of 'which provider' to 'which workflow'. When Cline documents that GLM cleans up dead code and verifies the build, while Opus leaves behind type errors that pass tests but break production, the argument for proprietary superiority is over. The fear of vendor lock-in following the access restrictions on frontier models is driving the rest. Anyone who opens their routing gateway now and uses open weights for standard tasks not only saves money but also regains control. This can be tested tomorrow morning, not after the next architecture offsite.

Jefferies Warns: Memory Prices Continue to Skyrocket

Jefferies Equity Research anticipates a price jump of 40 to 50 percent quarter-over-quarter for Q3 2026, followed by another 30 to 40 percent in Q4. For 2027, a 40 to 45 percent year-over-year increase is expected, with an initial correction in average prices not anticipated until 2028 at the earliest, when 15 to 20 percent new capacity comes online. The big three—Samsung, SK Hynix, and Micron—offer no relief, and the hope for cheap memory from China is proving to be a myth: CXMT and YMTC are selling at prices similar to everyone else; their advantage is volume, primarily for the domestic market. Currently, 50 percent of capacity is tied up in long-term contracts, with Micron alone having signed 16 Strategic Customer Agreements, and this share could rise to 70 percent. This leaves less and less for consumer products like PCs, laptops, consoles, and smartphones, with end-user prices rising across the board. Apple is already lobbying to bring CXMT on board as an additional source. The driver behind all this is the AI demand from hyperscalers, which is redistributing the market. → wccftech.com

Synthszr Take: The real leverage is in one number: 50 percent of memory capacity is already gone, tied up in long-term contracts with hyperscalers, and the target is 70 percent. Whoever buys AI compute first leaves the shelves emptier for everyone else, and that's exactly what ends up on the bill for every notebook and smartphone. The China hope was a pipe dream; we wrote about exploding token costs at the end of May, and now the pressure is moving one level deeper into the hardware. For product planning, this means the Bill of Materials for 2026 and 2027 needs to be calculated now, not when the purchasing department puts the first 50-percent invoice on the table in Q3. Anyone betting on memory-hungry on-device AI should secure supply commitments contractually while there's still something to secure. Apple isn't lobbying CXMT for fun; this is a preview of the procurement reality for the next two years. The good news lies in the discipline such a bottleneck forces: scarce resources turn featuritis back into real prioritization.

Santander Releases AI Governance Stack as Open Source

On June 21, Banco Santander became the first major bank in the world to release its complete AI governance stack as open source: 14 repositories on GitHub, all under Apache 2.0. It includes guardrail optimization, mechanical enforcement of decisions, fairness testing, and the generation of synthetic fraud graphs. The code is open for any competitor, FinTech, or regulator to fork. Neither JPMorgan, Goldman, nor HSBC has done anything comparable. Meanwhile, these same Wall Street firms are moving in the opposite direction: JPMorgan and Goldman Sachs have pulled Claude models from their employees' toolkits, effectively becoming a kind of border control for AI use in America's financial sector. Two banks, two philosophies, the same month. → Linas from Linas's Newsletter

Synthszr Take: Santander is giving away its compliance recipe for free, and that's strategically smarter than it seems at first glance. Whoever sets the standard by which everyone else is audited defines the rules of the game for the entire market. The EU AI Act forces every bank in Europe to do the same homework anyway, and Santander has now written it and posted it on the wall. When regulators and competitors fork the same code, Santander's logic becomes the de facto standard—a moat that looks like generosity. On the other side of the Atlantic, JPMorgan and Goldman are blocking a third-party model, which smells like caution in the short term and a loss of control over their own people in the medium term (they'll just use Claude privately). Openness as governance beats isolation as governance, because trust needs to be auditable, not hidden. Anyone still treating their AI guardrails as a corporate secret in 2026 is flogging a dead horse.

Code Crash: Anthropic is Now Hiring Product Managers for Better Intents

Anthropic has instructed its growth team to hire more product managers, not fewer. The reason: Claude Code has elevated the engineering organization to about three times its actual headcount, and the bottleneck has shifted from the development environment to the people who decide what should be built in the first place. The numbers back this up. New monthly questions on Stack Overflow have plummeted by about 77 percent since ChatGPT's launch in November 2022, Amazon's Kiro team compressed feature builds from two weeks to two days, and an AWS team completed a re-architecture originally planned for 30 developers over 18 months with 6 people in 76 days. The classic ratio of one PM to eight developers is now effectively closer to 1:20 because each developer delivers more per day. LinkedIn has replaced its associate PM track with a 'Product Builder' program that trains generalists across product, design, and engineering. According to the Stack Overflow Survey 2025, 84 percent of developers use AI tools, but 46 percent do not trust the output—a significant increase from the previous 31 percent. → Synthszr

Synthszr Take: This is exactly the point I described in Code Crash as the Product Engineer, now backed by hiring decisions instead of theories. The velocity is there, coding is solved, the bottleneck is one level up: with the intent. The person who can precisely describe what 'right' means is suddenly the most valuable person in the room, and that has little to do with prompting tricks. The fundamentals are becoming more important, not obsolete, because when a memory leak takes down production at 3 a.m., no agent can close that loop alone. The recursion is fascinating: Anthropic builds 80 percent of its codebase with its own tool and still hires more product minds because the machine produces features faster than the team can make decisions about them. Anyone restructuring their organization now shouldn't triple the engineering team, but rather triple the ability to ask good 'what' questions. This can be decided in the next quarterly plan, not after the offsite after next.

The Real Bottleneck is Verification

Gennaro Cuofano argues in his newsletter that the gap between perception and reality in the AI industry has never been wider. While the outside world talks about the bubble bursting and stagnating token consumption, the technology continues to accelerate beneath the surface. Token-maxxing is ending, token routing is becoming the norm, and a governance layer is inserting itself as its own structural layer on the AI map. Cuofano breaks down the familiar paradox (great demos, disappointing operations) into three symptoms of the same problem: agents that still fail at real tasks; users who don't leave the chat window; and a market that continues to fund classic software instead of agent-native replacement solutions. His core point: two curves are being confused. Peak Capability, meaning what the best model can do on its best day on a verifiable task, is racing ahead. Reliability at Scale, meaning what the typical model does in the long tail of fuzzy tasks, is barely moving. Software leads because it is the most verifiable domain in the world: code either compiles or it doesn't, and you can run a thousand identical copies in parallel. → The Business Engineer

Synthszr Take: Cuofano describes from a model perspective what I call the Interpretation Gap from an organizational perspective in Code Crash. Where the world provides a clean signal (it compiles, proves itself, wins the game), the model reliably learns. Where no one can quickly say whether a strategy was good or a contract was fair, it learns to be agreeable rather than correct. Right there, at every interface between proposal and decision, lies the real bottleneck, and no platform solves it for you. Gartner expects that around 30 percent of all GenAI projects will be discontinued after the proof-of-concept stage, almost never due to technology, but because verification and liability issues remain unresolved. Anyone investing money in the next model benchmark instead of in verifiable feedback loops for their own, fuzzy tasks is funding the wrong curve. The compute discipline of the next two years will be decided by whether you build your own answer key where the world doesn't provide one for free.

davaidavai 180

In issue 180 of his marketing link mix, Gerald Hensel curates the two major trends post-Cannes: On one side, the new super-pipelines from Meta & Co. that are taking over more and more parts of creative work. On the other side, a counter-model is forming, which he calls 'Human Premium.' 'Guaranteed without AI' is becoming a market position, according to the cited State-of-Brand-Report, a 'human-made premium' that AI-first brands by definition are not supposed to be able to match. Polaroid demonstratively positions itself 'on the beach against the data centers,' while other cases like Ikea Canada's World Cup flags made from its own product range or Oura's Ring campaign show that clever communication works without machine-slop. Alongside this are texts about the transformation of the creative brief from a persuasion tool to a memory tool and about the question of what an idea actually is. Hensel's conclusion: Who wins in man versus machine won't be decided in a Terminator-like fashion, but in a future of our own choosing. → Gerald Hensel

Synthszr Take: 'Guaranteed without AI' as a business model is an elegant capitulation, nicely packaged as a stance. Brands now betting on 'Human Premium' are turning their own inertia into a virtue, and the market will reward this for about as long as organic labels on yogurt cups last. The real game is being played elsewhere: Drucker said decades ago that a company has only two functions—marketing and innovation—and both mean customer-centricity. AI is the lever to scale exactly that, not the enemy you stand on the beach against. Polaroid is welcome to stay analog, but the interesting cases from this issue (Ikea building World Cup flags from its own shelves) show the better path: close to people, working with what's already there, and using the machine as an enabler. The industry should stop dwarfing itself down to advertising and instead seize the massive opportunity to make marketing a central control function again. Anyone who reads AI only as a threat has misunderstood the job.