Japan Launches Fable Alternative — EU Blocks It

Japan Counters Claude Mythos with Fugu — But Without Europe

Sakana, the increasingly enterprise-focused AI startup from ex-Google Brain's David Ha, launched Fugu last night: a multi-agent orchestration system that delivers frontier performance through a single, OpenAI-compatible API. Instead of relying on a single monolithic model, Fugu (Japanese for pufferfish) dynamically routes requests to an interchangeable pool of specialized agents. The timing is no coincidence: On June 12, following a U.S. export control order, Anthropic cut off public access to its top models, Claude Mythos 5 and Fable 5. Fugu surpasses them on benchmarks: 93.2 on LiveCodeBench versus Fable's 89.8 and 95.5 on GPQA-Diamond. The prices are steep: Fugu Ultra starts at $5 per million input tokens and $30 per million output tokens. The catch: Which models Fugu chooses internally remains proprietary and hidden from the user, and in the EU, the service isn't running at all for now due to open GDPR questions. → venturebeat.com

Synthszr Take: On June 12, Anthropic shut down Claude Mythos 5 and Fable 5 outside the US. Three days later, a system arrives from Tokyo that surpasses them on benchmarks. This should make us in Europe think, and for the right reason. Sakana has delivered what Mistral & Co couldn't. We talk about sovereignty, found consortia, write strategy papers. David Ha read the export order and built a finished product to fill the gap. Fugu scores 93.2 on LiveCodeBench versus Fable's 89.8. This isn't an announcement. It's live. And what does the EU do? It blocks the service. Open GDPR questions, Fugu isn't launching in Europe for now. We just witnessed live how quickly a US provider can turn off the tap. In the same moment, a friendly nation delivers the first real alternative, and Brussels turns it away at the border. That is strategically foolish.

Japan and Europe are in the same boat. Both depend on US models, neither has its own frontier stack, and both are watching the race between Washington and Beijing from the sidelines. The logical next move is clear: call instead of block. Common standards, a common market, a second pole not based in California. A smart alliance with Sakana would be the first real move Europe has made in this game. Fugu is no saint here. The routing logic remains a black box, the user doesn't see which model is working on their request, and $30 per million output tokens is steep for a layer that only coordinates. Sakana protects itself from vendor lock-in by becoming a gatekeeper itself. This is exactly what needs to be negotiated. But you can only negotiate if you're at the table. If you want sovereignty, you have to want it even when it's inconvenient. Otherwise, Europe will remain the market that watches others build the infrastructure it will depend on tomorrow.

Google's Talent Exodus: A Nobel Laureate Goes to Anthropic

John Jumper is leaving Google DeepMind after nine years to join Anthropic. Jumper co-developed AlphaFold, the AI for predicting protein structures, and received the Nobel Prize in Chemistry for it along with Demis Hassabis. His departure comes just days after Gemini co-lead Noam Shazeer moved to OpenAI, marking two prominent minds leaving in a single week. Jumper plans to “take a break” before starting at Anthropic, just in time for a science event scheduled for June 30. He reportedly also recently worked on coding tools for the enterprise sector at Google, a field where the company lags behind its frontier rivals. After Google's models led the pack in 2024 and 2025, 2026 feels like a step back compared to OpenAI and Anthropic. Science was considered DeepMind's home turf, and that's exactly where there's now a gap. → The Rundown AI

Synthszr Take: Talent is the real moat in the AI business, and Google is currently losing it to the two companies it's competing against. Jumper isn't just any researcher; with AlphaFold, he delivered the one flagship project that elevated DeepMind beyond mere language models. The fact that he's going to Anthropic, whose revenue we saw heading toward $20 billion in early March, says more about the gravitational pull of valuations than about Google itself. A corporation with this research depth can replicate models, but it can't pull a second Nobel laureate off the shelf. The sober question for Sundar Pichai is why the best minds prefer stock packages from a highly valued startup over the security of a trillion-dollar corporation. Anyone who answers with just more compute and more salary hasn't understood the question. It's about the feeling of being in the fastest car, and Google has squandered that feeling in twelve months.

Open-Weight Models: The 'ChatGPT Moment' for Your Pocket

For the first time, a freely downloadable model with an MIT license is playing on par with the most expensive closed systems. We're talking about GLM-5.2 from Zhipu AI (Z.ai), which is thrilling developers across the board: Matt Velloso, former VP at Meta, DeepMind, and Microsoft, calls it 'the first open-source model that qualifies as a daily driver.' Jeremy Howard, co-founder of Fast AI, puts it on the same level as Opus 4.8 and GPT-5.5. The numbers back up the hype: a one-million-token context window that holds an entire codebase at once, first place on independent open-weight leaderboards, and the top spot for front-end coding on Arena AI. It all runs on your own hardware, for free, and under a permissive license. We already noted in mid-June how GLM-5.2 overtook GPT-5.5 at a fraction of the price; now, the gap to the closed-source camp has practically disappeared. → Linas from Linas's Newsletter

Synthszr Take: At the end of May, Anthropic was the undisputed number one with a record valuation; now, an open-weights model from China is delivering Opus-4.8-level performance at a fraction of the operating costs. This is the Jevons paradox in action: the cheaper inference becomes, the more of it is consumed, and value shifts from model weights to orchestration. Sakana's Fugu Ultra makes this visible, because it's not one giant model that wins, but the smart coordination of specialized models behind an API. In the OH-SO-Vendor-Maps, the sentence has stood since round six that the choice of orchestration framework has a longer half-life than the choice of model, and GLM-5.2 confirms this once again. Anyone who ties their coding stack exclusively to a proprietary provider today is building a lock-in risk that can become very expensive in six months. 'Reasonable Sovereignty' means: a frontier model as the standard, an open-weights backup like GLM-5.2 or Llama alongside it, and the logic in an agnostic framework like LangGraph. This architecture can be configured in the pilot phase, not after the next strategy offsite.

GLM-5.2 Just Ahead of Kimi K2.7: Two Chinese Open-Source Models Reset Frontier Prices

Zhipu AI's GLM-5.2 and Moonshot's Kimi K2.7 Code, both launched within days of each other in mid-June, have now been tested head-to-head by five independent reviewers, and the early verdict puts GLM-5.2 slightly ahead. For quick one-shot tasks, the two traded victories, with Kimi often being faster and adding features without being asked. The difference became apparent on a second look: GLM's builds held up better to code inspection, while Kimi's polished surfaces hid more bugs. GLM-5.2 runs on 744 billion parameters (40 billion active), has an MIT license, a one-million-token context window, leads Artificial Analysis' Intelligence Index with 51 points, and beat GPT-5.5 on GDPval. Kimi K2.7 Code is the leaner instrument: one trillion parameters, 32 billion active, is the only one of the two that can read images, but its context window of around 256,000 tokens was called tight for production code by several reviewers. For a Three.js racing game, GLM needed one prompt and 40,000 tokens, while Kimi needed a follow-up and 110,000. And the price: GLM comes in at about 50 cents per task, while Kimi starts at $15 a month. → Marcus Schuler

Synthszr Take: Five subjective individual runs are not a benchmark; the pattern across all of them counts for more than any single test. Still, the direction is clear: two freely available models from China are playing in the same league as Opus 4.5, and one of them completes a coding task for about 50 cents. In our vendor map, we recommended Claude Code and Cursor as the standard stack, with Cline plus an on-premise model for regulated workloads. This third slot is exactly what's becoming interesting, because GLM-5.2 under an MIT license is what we called 'Reasonable Sovereignty': a multi-region and open-source backup instead of hyperscaler lock-in. Anyone who sets up their architecture to be model-agnostic in 2026 will have a choice in 2027 instead of being dependent on Anthropic. Don't overestimate GLM's code-reading ability yet; the reviewers judged by eye and had Claude double-check. But ignoring the price would be the most expensive mistake: when a task drops from $15 to 50 cents, it's not the tooling that shifts, but the math behind it.

Getty Images: OpenAI Partnership Sends Stock Soaring 124 Percent

Getty Images has announced a multi-year licensing and product partnership with OpenAI, and its stock shot up 124 percent in early trading. OpenAI will license Getty's library of images, videos, and metadata for training and improving its models, while Getty, in return, will integrate OpenAI's generative tools into its own products. The deal comes at the right time: in the first quarter, Getty's revenue was $226.6 million, a 2.5 percent decrease year-over-year, and its traditional licensing business is threatened by tools that generate stock photos in seconds. Getty was one of the first major content providers to take legal action, suing Stability AI in 2023 for scraping millions of copyrighted images. Before the announcement, the stock had been trading below one dollar for months. The OpenAI agreement follows a licensing deal with Perplexity from 2025. Investors are now waiting to see if Getty can turn its curated library into a more valuable AI asset. → Techpresso

Synthszr Take: Two years ago, Getty was still suing AI companies; now, the company is licensing its library to the very same people. This isn't a change of heart out of conviction; this is a stock that was trading below a dollar for months and is grasping at straws. The point the 124 percent jump shows is this: Getty's value no longer lies in selling individual stock photos, but in the curated, cleanly tagged, and legally sound dataset behind them. This is exactly what OpenAI needs for training and what no one can scrape with impunity anymore, now that the courts have woken up. The intriguing question is whether $226 million in quarterly revenue is enough to build recurring licensing revenue from this asset, or if OpenAI will have the data after training and Getty will just be celebrating a one-time effect. Anyone hoarding curated content should start pricing it as an AI training asset now, while the model builders are still paying. Mediocrity is disappearing here too; what remains is what is truly unique and hard to replicate.

China Strikes Back: Export Controls and Procurement Bans for US Firms

On Monday, China imposed new restrictions on exports and public procurement against dozens of US companies. According to Nikkei, the measures primarily affect firms in the drone, rare earths, and related technology sectors. These actions come despite Trump's visit to Beijing in May, where both sides spoke of de-escalation. Since then, the trade measures have escalated on both sides, and Beijing is now using two levers at once: access for Chinese suppliers and access to its own procurement market. This is part of an escalation that began with the draconian US export controls against Anthropic in June, which we wrote about at the time. Those who depend on rare earths will feel it first. → Marcus Schuler

Synthszr Take: The real message lies in the rare earths, not the drone list. China controls about 90 percent of the processing, and this lever determines who can build AI hardware and sensors. Washington plays export control, Beijing plays import dependence, and both forget that in a hyper-competition, the loser is often the third party: Europe, which controls neither the models nor the raw materials. We are not in the lead; we are on our backs, and in this position, a strategy offsite won't help, but an honest inventory of our own supply chains will. The question is both banal and brutal: which critical inputs depend on a single source, and what happens the day that source is shut down? That can be calculated tomorrow morning. Anyone who waits until the delivery fails to arrive has chosen the wrong moment.

AI Code Generation: Standardization in the Development Team

According to a survey of 219 engineering leaders, 48 percent of code is now generated by AI, but organizations are lagging brutally behind: only 19 companies have adjusted role descriptions, just 15 have updated onboarding for new people, and 32 have revised productivity measurement. 55 percent of leaders fear their teams are losing a shared understanding of the codebase, and 39 percent doubt they can still ship safely with increasingly agent-written code. The core problem isn't bad AI code, but a lack of standards: one agent uses Jest, another Mocha; one uses async/await, another Promises. Everything works, but the codebase becomes harder to understand with every commit. IBM calls the result 'intent debt': the goals and guardrails that should tell both humans and agents what 'good' means are still written for the old world. The proposed solution is project-level rules, i.e., explicit conventions that give AI agents the same framework that human developers would follow. → The Deep View

Synthszr Take: This is exactly what I described in 'Code Crash,' just under a different name. Implicit knowledge has always been the most expensive anti-pattern, and with agents, the bill comes due immediately. An experienced developer knows the unwritten rule, the bug from 2019, the module you don't touch. An LLM only sees what's in the code, and if the code hides its own structure, the agent optimizes for the wrong things. Therefore, project-level rules are not bureaucratic overhead, but knowledge archaeology in its purest form: spend one hour a week making an implicit rule explicit, and then the agent reads the same convention as a junior developer. Anyone who has 48 percent of their code written by machines but keeps the role descriptions from the old world is building its own complexity trap. These standards can be anchored in the repo tomorrow morning, not after the next strategy offsite.

JD.com: 700,000 Couriers to Be Replaced by Robots

Richard Liu, CEO of Chinese e-commerce giant JD.com, openly stated at the APEC China CEO Forum in Shenzhen what most tech bosses avoid: robots will “sooner or later” make the company's 700,000 delivery workers redundant. At the same time, he promises not to lay off a single frontline employee. His program is called 'Nirvana': JD has signed contracts with around 120 schools across China to retrain couriers as maintenance engineers and AI trainers. JD builds the robots itself, operates unmanned warehouses, drone delivery, and self-driving delivery vans, and is already testing airport robots in Shenzhen. The backdrop is tense: China has around 320 million gig workers this year, about 40 percent of urban employment, and youth unemployment was at 16.3 percent in April. Beijing now tracks the impact of AI on jobs as a national priority. The open question remains whether 120 schools can retrain a workforce of 700,000 faster than automation overtakes them. → Techpresso

Synthszr Take: Liu's honesty is rarer than the technology he's announcing. While Western CEOs soften their message depending on the political winds (Bezos says one thing today, Altman another tomorrow), the JD chief lays the bill openly on the table and adds a promise on top that is mathematically almost impossible to keep. There won't be nearly as many robot maintenance roles as there are courier jobs, and he knows it. Nevertheless, the sequencing argument is smarter than the usual silence: whoever names the disruption early can at least get retraining started, instead of letting 700,000 people drive headfirst into a wall. We wrote in mid-March about the end of wage arbitrage when 100,000 Indian IT jobs disappeared; here, the same thing is happening one level down, with blue-collar jobs, and the speed is more brutal. The honest version beats the sugar-coated one: JD's own employment data over the next few years will show whether 'Nirvana' is a plan or a PowerPoint illusion. Any government or company not investing in real retraining now is flogging a dead horse.

Data Centers and Electricity Prices: A Causal Analysis for the USA

A new paper by John Bistline (arXiv, June 2026) flips the common narrative: data centers did not drive up average electricity prices for US retail customers between 2015 and 2024, but moderately lowered them. The authors use an instrumental variables approach to cleanly separate the causal effect from other factors. The logic behind this is sober economics: power systems have high fixed costs, transmission and distribution scale, and the unit costs of generation decrease. Sustained growing demand distributes these fixed costs over more kilowatt-hours, thus lowering the average price. The study finds these economies of scale across generation, transmission, and distribution, as well as between different customer classes. The limitation is clearly stated in the abstract: future supply constraints could reverse the effect. → Techpresso

Synthszr Take: Here, someone is arguing with data against a gut feeling that has permeated every article on AI and power for months. The assumption that hyperscalers make electricity more expensive for everyone sounds plausible and is simply proven false for the period 2015 to 2024. The mechanism is the Jevons paradox in its unsexy version: more load on a fixed-cost-heavy system lowers the average price per unit, as long as supply keeps up. This is exactly where it gets interesting, as the authors set their own guardrail: if grid expansion lags behind demand, the sign flips. For energy policy, this means the lever lies in expanding transmission and generation, not in slowing down data centers. Those who take economies of scale seriously will expand the grids faster and reap the digitalization dividend, instead of giving it away out of fear. This can be influenced by the speed of approvals and investment planning, long before the next bottleneck disproves the study.