Lunching with Trump Pays Off: Fable 5 Ban Could Collapse Like a Soufflé

Trump's Whims: One Lunch with Dario Amodei and Everything's Just Fine

In an Axios interview published on June 19, Donald Trump stated he no longer views Anthropic as a national security threat, two days after meeting with CEO Dario Amodei at the G7 summit in Évian-les-Bains, France. During a working lunch on June 17, Amodei, along with Demis Hassabis, had proposed a US-led democracy coalition for AI standards to about a dozen tech CEOs, deliberately excluding China. According to Trump, the trigger for the whole escalation was a competitor who is also a co-owner: This fits the description of Amazon, which has invested $8 billion in Anthropic and runs competing models on AWS Bedrock. Amazon reported a vulnerability in Mythos, whereupon Commerce Secretary Lutnick sent out the export control order at 5:21 PM on June 12, giving Anthropic 90 minutes to take Mythos 5 and Fable 5 offline. The order and the Pentagon's classification from March 3 formally remain in effect. The timing hits Anthropic hard: a $965 billion post-money valuation after the $65 billion Series H-1, S-1 confidentially filed since June 1, with a Nasdaq listing planned for October 2026 with a volume of over $60 billion. Alex Stamos and nearly 150 security leaders called the authorities' reaction excessive. → Techpresso

Synthszr Take: A security policy that can be reversed in a week by a nice lunch on the sidelines of the G7 is not policy, it's a whim. Six days ago, Anthropic was a national threat; now, a pat on the back for Amodei is enough for the winds to change. The real scandal is in the footnote: Amazon, having invested billions in Anthropic, snitched on its own partner to the government. This is what a strategic partner looks like—one that would rather slow down a competitor via government action than win in the market. For anyone building on US AI infrastructure, this is the uncomfortable lesson: The rules of the game depend on a phone call between a CEO and the president, not on reliable laws. Anyone basing their product strategy on that is building on quicksand. From now on, second suppliers, open-source options, and European compute capacity are not nice-to-haves, but standard equipment.

Fable Ban: Open-Source Models Are More Than a Stopgap

An internal ban on Fable is forcing development teams back to open-source models they can run on their own metal. The triggers are the usual concerns: Who owns the data, where does the inference run, and how reliable is a proprietary system that can change the rules at any time? Instead of tying themselves to a single provider, teams are turning to freely available models that can be adapted to their own needs. The New Stack describes this as a stopgap measure, as open models still don't quite keep up in some disciplines. This gives providers like Ada and Anthropic, who rely on closed systems, one less argument. The trend fits the picture of a market seeking trustworthy and controllable AI systems, while computing resources and geopolitics reshuffle the deck. → The New Stack

Synthszr Take: 'Stopgap' is the interesting word here, because what starts as a temporary fix today will become the standard architecture tomorrow. This was exactly Zuckerberg's calculation when Meta made Llama freely available, thereby drying up the markup for a good model. Anyone who reads a ban on Fable as a setback is looking in the rearview mirror. The real movement is in the other direction: Teams running their models on their own metal are regaining control over data flow, costs, and availability instead of delegating it to a gatekeeper. The gap between open and closed has been shrinking for months, and any provider whose business model is built on an inference surcharge should be watching nervously. Those who now switch their AI strategy to interchangeable, self-controllable models are buying themselves room to maneuver that no proprietary license model can take back. This can be started in the next sprint, not after the strategy offsite after next.

The Economics of China's Cut-Price Models

DeepSeek launched V3 at $0.27 per million input tokens, while GPT-4o was charging $2.50 at the time, about nine times as much. By mid-2026, the gap has widened even further: Chinese frontier models are priced at $0.10 to $0.50, while Western ones are at $2.00 to $15.00. Moonshot's model is asking $0.06, Zhipu's GLM-4.7 beats Claude Sonnet on the MATH benchmarks, and DeepSeek V4 is within 3 percent of GPT-4o on MMLU-Pro. The main reason is architecture: Mixture-of-Experts activates only 37 out of 671 billion parameters per token for DeepSeek, which is 2.2 percent of the FLOPs of a dense model. Added to this are FP8 training, a curated 14.8 trillion tokens instead of 40 trillion via brute-force, and multi-token prediction. On the hardware side, an H20 hour in China costs $0.80 to $1.20 with identical memory bandwidth to the H100, which costs $3 to $5 on AWS. And a senior ML researcher in Hangzhou costs $100,000 to $250,000 instead of $400,000 to $700,000 in San Francisco. → Synthszr

Synthszr Take: The most exciting sentence in the whole text is that the Chinese labs are still profitable. This isn't a price-dumping story, but real compute discipline built into every layer of the stack. And this is precisely the message for everyone here who wants to build their domain-specific software: The execution power that cost 100 million two years ago is now available for a fiftieth of the price. This is the Jevons paradox in its purest form; cheaper inference opens up hundreds of niche markets that were too small yesterday. Anyone in Beijing computing with 37 billion parameters instead of 1.7 trillion has understood that capacity and inference costs are two different things, and every product manager looking at their token bill should internalize that. The moat for these models isn't skill, which has become interchangeable; the moat is the data you don't give away. The models can come from California or Hangzhou—that can be a vendor decision made tomorrow morning; sovereignty is decided elsewhere, by your own domain data.

One API for 50+ Chinese Models: Access to the Price Revolution

AIWave provides a translation layer between your own application and eight Chinese labs, making more than 50 models accessible through a single OpenAI-compatible endpoint. The numbers that make this interesting: DeepSeek V4 costs about 3% per token of what GPT-4o costs, GLM-4 runs at about one-twentieth of the price on comparable benchmarks, and Qwen delivers multilingual performance on par with Claude for a rounding error on the cloud bill. The real problem so far has been friction: eight providers, seven authentication schemes, four API formats, Chinese-language dashboards, and separate wallets. AIWave normalizes this down to two changed strings in the code, base_url and model in the format provider/model-name, while the rest (streaming, function calling, temperature) works as usual. The catalog ranges from deepseek/deepseek-v4-pro with 128K context for reasoning, to qwen/qwen-3-max with 256K for long documents, to budget models like yi/yi-lightning. Built-in features include load balancing and fallback chains, for example, from DeepSeek to Kimi upon hitting rate limits. Data as of: June 2026. → Synthszr

Synthszr Take: The economics were already absurd in 2025, but nobody was using them. That's precisely the point: it's not the model that decides, but the friction in accessing it. As long as a team has to struggle through DashScope-HMAC, expiring JWTs, and ByteDance-SigV4, they'll stick with the one Western provider and pay the premium, even if the spreadsheet says there's a 30x difference. AIWave isn't selling a model; it's selling the disappearance of integration work, and that's the more expensive half of the bill. Anyone who properly sets up cost telemetry per use case in my Code Crash vendor map can now offload nightly refactorings or bulk summarizations to a third-of-a-permille model and reserve the frontier model for the hard reasoning cases. The Jevons paradox applies immediately: cheaper tokens don't mean lower spending, but ten times more calls. In February, we wrote that what happens in China doesn't stay in China. Now it's on your desk in two lines of code, and you can test it this week, not after the next architecture review.

Universal Jailbreaks: When AI Safety Nets Tear

Researchers have shown that attacks on large language models can be constructed fully automatically. You append a specially computed string to a user query, and the model obeys commands it should refuse due to fine-tuning. Unlike manually crafted jailbreaks, which providers quickly patch, these attacks can be generated in any number. They were optimized on open-source models whose weights are public, but the strings transfer to closed systems like ChatGPT, Bard, and Claude. Particularly unsettling: The authors consider it an open question whether the problem can be fixed completely at all. Similar adversarial attacks have plagued computer vision for ten years without a solution. The researchers disclosed their findings to the affected providers beforehand, which is why individual strings are likely ineffective by now. → Simon Willison from Simon Willison's Newsletter

Synthszr Take: The uncomfortable news is in the subordinate clause: This could be inherent in the nature of deep neural networks and thus permanently unsolvable. As long as an LLM is just spitting out text, it's a PR problem. Once it acts autonomously—sending emails, executing code, reacting to web searches—the harmful output becomes an open door for anyone who appends the right character string. In May, we wrote about Anthropic's TACO Day and the Pentagon access, and this is exactly where two lines collide: you sell autonomy to security-critical customers while being unable to guarantee that the guardrails will hold. The compute discipline of the coming years, therefore, lies not in a bigger model, but in the architecture around it: What is an LLM allowed to initiate without human confirmation? Whoever answers this question today, instead of after the next incident, is building the real moat before someone writes a consultant deck about it.

GPT-5.6 Is Likely Coming Next Week

OpenAI is launching GPT-5.6 next week, presumably including Mini and Pro variants. The context window is expanding to 1.5 million tokens, accompanied by better coding over long periods and faster Codex response times. The prices are set to undercut Anthropic, which is helpful right now as regulatory issues in the US are affecting the availability of Claude Fable 5. In parallel, Yann LeCun has spoken out: he calls Musk's xAI a 'failure' that can't keep up with OpenAI and Anthropic, and warns of a 'big bubble explosion' if the labs don't raise their prices or cut costs. The operational costs of frontier models remain the unresolved issue behind all these releases. → TLDR AI

Synthszr Take: 1.5 million tokens of context sounds like a leap, but in the daily life of engineering teams, it's primarily a bill. More context means more compute per request, and this is where the price war against Anthropic gets interesting: OpenAI is subsidizing market share while LeCun talks about a bubble explosion. Both can be true at the same time. Those already running Codex as a workspace standard (we noted it in the vendor map with workspace lock-in) will get faster response times and better refactorings with 5.6, but they'll pay the price in lock-in, not just in dollars. The interesting question for the coming months is not the benchmark delta, but whether inference costs will ever be reflected in the prices or will be permanently buffered by investor money. As long as that remains an open question, the same sober advice applies to every team: models change quickly, but the orchestration framework underneath has a longer half-life. Build on that, not on the version number.

ChatGPT Becomes a Personal Assistant with a Schedule Book

OpenAI has given ChatGPT a new 'Scheduled' page in the sidebar that bundles all active tasks in one place. From there, they can be viewed, paused, edited, or deleted. Research tasks search the web and connected apps and only report back when something actually changes. Everything can be scheduled for specific times or parts of the day (morning, noon, evening), with a task running at most once per hour and automatically pausing during inactivity. The feature is available for Plus, Pro, Business, and Enterprise users, with different limits per plan. The older 'Pulse' feature is being discontinued and merged into scheduled tasks. → Techpresso

Synthszr Take: Here, the chatbot gets a memory for time, and that shifts its role. As long as ChatGPT only answers when you ask, it remains a tool you operate. With scheduled tasks that run in the background and only ping you when there's a real change, it becomes an operating stream that you direct rather than operate. This is precisely the move from 'human in the loop' to 'human in the lead': not reacting when the machine reports a problem, but setting a direction and letting the machine work. The maximum frequency of once per hour and the pausing during inactivity show that OpenAI is disciplining its compute, as these background runs cost processing time without a direct user click. Anyone who wants to test how this feels tomorrow morning should set up three recurring tasks and see what's left after a week. It wouldn't be my tip, but the finding is clear: The assistant that starts on its own is closer than most people think.

Vibe-Coded Apps: Security for the New Software Generation

Retool is repositioning itself and turning the dirty secret of the vibe-coding wave into a selling point: 'Secure your vibe-coded apps'. The platform promises internal software that is enterprise-grade from day one—meaning no subsequent rebuilds, no audit stress, no IT veto. According to the company, more than 10,000 teams use Retool for production-ready AI applications. The references are solid: Ramp reports $8 million in savings and over 20,000 hours saved, DoorDash reports $6 million and 36,000 hours, and Orangetheory a tenfold reduction in development time across 1,600 studios. The promise is: business teams build fast, IT retains full visibility, and governance remains centralized. In addition, Retool is delivering a new React AI app builder and a 'State of AI Governance 2026' report. → TAAFT - There's An AI For That

Synthszr Take: In spring 2025, the diagnosis was still alarming: Veracode stated that 45 percent of AI-generated code failed security tests. Now, Retool is selling this exact gap as a product category, and that's the most honest move in the whole vibe-coding circus. The interesting shift is in the details: security is moving from the person who writes the code (or doesn't understand it) to the platform that delivers it. Feature flags, automated security gates, audit trails for every prompt: in early 2025, this was still the Wild West; today, it's a compliance feature with a price tag. Any hidden champion looking to build its internal tools no longer has to ask the old question of whether to be fast or secure. Both can be decided tomorrow morning, not after the next strategy offsite. The point is: the toolchain has matured, and that's precisely why a risk is turning into a business model.

Europe and AI: A Doomsday Scenario as a Wake-Up Call

A Brussels-based think tank has published a thought experiment called 'Europe 2031' in which the EU is crushed between the US and China: America builds data centers and fires people, China builds robots, and Europe takes long lunch breaks and leaves administrative work to Claude. The timing was fortunate, as one day after publication, the Trump administration blocked 'foreign nationals' from accessing an Anthropic model named Fable, briefly making one of the central predictions a reality. The scenario went viral during the G7 week, was read by Members of the European Parliament, and cited in British-German Track 1.5 talks. Author Maximilian Negele, formerly of the Rand Corporation, speaks of a 'slow-moving car crash' and a translation barrier between Brussels and San Francisco. In the plot, the US monopolizes 70 percent of the world's compute capacity, while the EU desperately tries to use the Dutch company ASML as a last resort. Skeptics point out that several of the mentioned megadeals have long since fallen through: the $100 billion pact between OpenAI and Nvidia vanished into thin air in February, the $300 billion with Oracle is considered dubious, and the bulldozers in Texas are at a standstill. The authors remain calm and concede that one or two AI companies might go bankrupt. → www.theguardian.com

Synthszr Take: The scenario works as a scare tactic, and that's precisely why it's a bit suspicious. When the most impressive figures in the plot ($100B OpenAI/Nvidia, $300B Oracle, the Texas bulldozers) have already been debunked in reality, then the whole doomsday edifice is built on sand that is currently slipping away. The more honest story is taking place in Stockholm and Paris: Legora cracked $100 million in annual revenue in under eighteen months and supplies a fifth of the highest-grossing US law firms, and Yann LeCun's AMI Labs raised the largest seed round Europe has ever seen with $1.03 billion. Europe's problem was never the depth of its research, but its translation into breadth, and AI is now delivering that. Last week's Fable ban (see 'The Anthropic Quake') clearly shows the cost of depending on US models—that's the real argument for sovereignty, not the spyware affair fantasy at the end of the paper. Fear mobilizes in the short term, but it doesn't build data centers. What Europe needs is compute discipline and the courage to finally give its hidden champions the capital to scale, instead of being spooked by a Brussels tragedy in five acts.

The Thermodynamics of Intelligence: A New Metric?

A preprint by Ishanu Chattopadhyay (arXiv, submitted in June 2026) attempts to make intelligence measurable as a physical quantity. The thesis: Intelligence is the lawful amplification of rare but permissible futures. A system is considered intelligent if it increases the probability of outcomes that would be unlikely under passive dynamics but remain within the rules of the respective field. For the authors, this necessarily implies recursive self-simulation: because the system is part of the world it models, it must include itself and its actions in the simulated futures. The central results are a necessity and a near-sufficiency statement that couple this architecture to a precise thermodynamic measure. The result is a universal scale that ranges from passive matter, through control loops and large language models, to humans as text generators and Maxwell's demons. → Techpresso

Synthszr Take: Behind the dense physics vocabulary lies an idea that is already being practiced in everyday product development. Anyone building a system that finds the rare, valuable, yet still permissible next step out of millions of possibilities is building exactly this amplification machine. The claim of a continuous scale is exciting: the same metric for a thermostat and for GPT forces honesty about what a model truly achieves. Whether the math holds up will have to be verified by other labs, and a single preprint without replication is, for now, a strong assertion. But the practical core can be tested immediately: does your system measure the accuracy of its own predictions about rare outcomes, or does it only optimize for the average case? Anyone who expresses intelligence in the future as lift over a random baseline instead of in vague benchmark points gets a metric that fits into a customer health score just as well as into a research paper. A metric that applies from a stone to a language model is exactly the kind of discipline the overheated debate is currently lacking.