๐ Washington freezes all foreign access to Anthropic's Fable 5 and Mythos 5
๐งฌ Nobel laureate John Jumper leaves DeepMind for Anthropic
๐ก Sakana's Fugu turns rival AI models into one team
๐ธ Wall Street's newest hire is an AI agent
โจ And more AI goodnessโฆ
โก The Signal
The US government just treated a commercial AI model like a weapon, and the fallout shows how strategic these systems have become. On June 12 the Trump administration ordered Anthropic to block every foreign national, including its own staff, from using Fable 5 and Mythos 5, and the company shut both down worldwide rather than break the order. The official trigger was a disputed jailbreak, but the real stakes surfaced a day earlier, when a US spy chief told the Senate that Mythos broke into almost all of America's classified systems in hours, not weeks. The rest of the issue rhymes with that: Anthropic poaches a Nobel laureate for AI-for-science, a fresh pricing chart shows the frontier still costs a fortune, and Wall Street quietly hands real work to AI agents. The throughline is control, over who gets to run the most capable models, and who profits when they replace human labor.
All the best,
Kim Isenberg
Demis Hassabis (left) and John Jumper, who shared the 2024 Nobel Prize in Chemistry (Getty Images via TechCrunch)
๐งฌ Nobel Laureate John Jumper Leaves DeepMind for Anthropic
John Jumper, co-creator of AlphaFold and a 2024 Nobel laureate in chemistry, is leaving Google DeepMind after nearly nine years to join Anthropic. Jumper led the team behind AlphaFold, the system that predicts protein structures and has now mapped more than 200 million of them, the work that won him and DeepMind CEO Demis Hassabis half of the 2024 Nobel Prize in Chemistry. On X he thanked Hassabis for "a real chance letting me lead the AlphaFold team just six months after finishing my PhD," and called DeepMind "a special place." His exit lands days after transformer co-inventor Noam Shazeer left Google for OpenAI, and points to Anthropic's growing bet on AI for science.
๐ tl;dr: Anthropic just landed one of the most decorated scientists in AI, and Google's research crown jewel keeps losing its stars.
MIT economist Christian Catalini, who coined the โNadella testโ (coindesk.com)
๐ผ Christian Catalini's 'Nadella Test': What's Left When the Model Is Pulled?
As frontier models get cheaper and nearly interchangeable, the question that matters is what value remains when you yank the model out. In a Forbes essay, MIT economist Christian Catalini names this "Nadella's test," after the Microsoft CEO's warning against drifting toward "a small number of AI systems capturing all the economic returns, while entire industries find their knowledge commoditized." Nadella's own answer: the edge is "not in picking the best model but instead in building a learning loop on top of models where human capital and token capital compound." Catalini's economics agrees: once automated execution becomes a commodity, durable advantage shifts to proprietary data, judgment, and accountability.
๐ tl;dr: When the model is a commodity, your moat is the data and judgment wrapped around it, not the model itself.
๐ก Sakana's Fugu Makes Rival AI Models Work as One Team
Sakana AI has launched Fugu, a single API that orchestrates several frontier models into one problem-solving team. Built on two ICLR 2026 papers, TRINITY and Conductor, Fugu uses a lightweight learned coordinator to assign models Thinker, Worker, and Verifier roles and to discover their own communication strategies, instead of leaning on hand-coded workflows. Sakana says Fugu and the higher-quality Fugu Ultra match or beat the strongest publicly available models, with Fugu Ultra topping SWE-Bench Pro at 73.7 and both clearing about 93 on LiveCodeBench. It even claims the pair stand "shoulder-to-shoulder with Fable 5 and Mythos Preview," Anthropic's most capable models, though Sakana concedes those two are not publicly accessible and so were not tested head-to-head. Demos run from autonomous paper reproduction to blindfold chess; API pricing starts in July.
๐ tl;dr: A coordinated team of cheaper models is now trading blows with the single strongest frontier system.
Today's Graph shows the smartest model can cost up to 38x the price of a cheaper rival that scores only 10 to 15 points lower. Before you reach for the most powerful model by reflex, run a quick cost-versus-need check on your own AI usage.
Why it helps: Most everyday tasks (drafting, summarizing, classifying, reformatting) do not need a frontier model. Matching each task to the cheapest model that clears the bar is the practical version of the "learning loop on top of models" Satya Nadella describes, and it can cut your AI bill by an order of magnitude.
Try this: "Here is a list of the AI tasks I run regularly: [paste]. For each one, tell me the minimum model tier that would reliably do the job (small and cheap, mid, or frontier), the failure mode if I under-spec it, and a one-line test I can run to confirm the cheaper model is good enough. Then flag the two or three tasks where paying for a frontier model is actually worth it."
๐ฌ Watch This
Trump talks about presidential power as if it has no real ceiling - and that alone makes the conversation revealing. Between Iran, China, Anthropic and Americaโs AI race, the interview shows how he connects military force, executive authority and technological dominance into one worldview.
The 3D AI tool that finally gives you geometry you can actually use
Every AI image-to-3D tool promises to change the game. Most deliver a broken mesh you spend hours cleaning before it touches your pipeline. Hi3D is different in one concrete way: it thinks aboutย theย geometry first.
The tool just opened access, and after hands-on time with it, what stands out isn't the speed โ it's the output. Models come out with usable geometry: clean edge loops, consistent polygon density, real topology. For 3D printing, no pre-flight repair. For game and animation pipelines, you can actually rig it.
The fine-detail surface preservation is what makers and concept artists are talking about โ textures that hold up at both render distance and print scale. That dual fidelity is rare at this price point. If you've burned hours post-processing broken meshes from other AI tools, this is worth your next 10 minutes.
Anniversary Offer: Up to 80% off + BOGO credit packs (limited time)
Industry leaker Leo, who has called several Anthropic releases correctly before, says Claude Sonnet 5 could ship as soon as next week, after the slug โclaude-sonnet-5โ quietly surfaced on a partner provider's API.
Anthropic may already have its next flagship in hand. The usually well-informed Andrew Curran posted that "A new, more capable version of Mythos has emerged from training. I don't know whether it will be called Mythos 5.1 or Mythos 6, or if Anthropic will keep it internal to accelerate further development, but it has arrived." Treat it as an unconfirmed report rather than an announcement: Anthropic has not commented, and the timing is striking, arriving the same week the government forced Fable 5 and Mythos 5 offline, which would make any quiet, internal-only successor especially hard to read.
Washington Froze Anthropic's Top Models. The NSA Just Revealed How Powerful They Are.
The Takeaway
๐ On June 12, the Trump administration ordered Anthropic to block every foreign national, including its own non-citizen staff, from using Fable 5 and Mythos 5; Anthropic disabled both worldwide to comply.
๐ The trigger was a disputed jailbreak of Fable 5 that Amazon flagged; White House adviser David Sacks said the government told Anthropic to "fix the jailbreak or de-deploy," and that Dario Amodei refused.
๐ A day earlier, Sen. Mark Warner said NSA and Cyber Command chief Gen. Joshua Rudd had told him Mythos "broke into almost all of our classified systems, not in weeks, but in hours."
๐ The cutoff spares no one, not even the Five Eyes allies or Britain's AI Security Institute, drawing comparisons to Cold War controls on nuclear and encryption technology.
For the second time in four months, the Trump administration dropped a Friday-afternoon bombshell on Anthropic, and this time it pulled the company's most powerful models offline entirely. Citing national security, the government issued an export-control directive ordering Anthropic to stop any non-American, including its own foreign-national employees, from accessing Fable 5 and Mythos 5. Rather than wall off part of its own workforce, Anthropic suspended both models for everyone. In its statement the company said it "disagree[s] that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people," and that it is "working to restore access as soon as possible."
Anthropic CEO Dario Amodei, who has publicly called for binding AI regulation (Getty Images via Fortune)
Both sides tell different stories about what set it off. Amazon, a major Anthropic backer, reportedly told the administration it had found a way to "jailbreak" Fable 5 and reach the dangerous cyber capabilities its guardrails were meant to contain. White House adviser David Sacks said officials asked Anthropic to "fix the jailbreak or de-deploy," and that Amodei refused; Anthropic counters that the flaw was narrow, already known, and no worse than what rival models expose. Commerce Secretary Howard Lutnick sent Amodei the letter spelling out the restrictions. Many read the move as less about one jailbreak than about leverage over a lab President Trump once branded an "out-of-control Radical Left AI company," especially after Pentagon chief Pete Hegseth boasted he had "kicked Anthropic out" of the Pentagon months earlier.
It is the second time since February that the Trump administration has moved against Anthropic (Alex Wong/Getty Images via TechCrunch)
The detail that reframes the whole episode landed a day before the order. On June 11, Sen. Mark Warner, vice-chair of the Senate Intelligence Committee, said Gen. Joshua Rudd, who runs the NSA and the Pentagon's Cyber Command, had told him that Mythos "broke into almost all of our classified systems, not in weeks, but in hours." That single sentence recasts the cutoff: the government is treating a frontier model like a strategic weapon, and that is what makes the episode bigger than a squabble over a chatbot. It is why the suspension exempts no one, not even the Five Eyes intelligence partners or Britain's AI Security Institute, the world's leading model-testing body, and why observers reach for Cold War analogies, from 1990s encryption export controls to the 1946 law that cut even close allies off from American nuclear research.
Allies who had spent months securing Mythos access for agencies, banks, and companies lost it overnight, putting close partners in the same bracket as Russia, China, and Iran. Anthropic says it is working to restore access and that spy agencies are already negotiating their way back, but Helen Toner of Georgetown's CSET warns the policy is "essentially equivalent to preventing any company affected from doing any further AI R&D work," given how many foreign nationals staff American labs. The same week, the administration told the Center for AI Standards and Innovation, the body that vets frontier models for dangerous capabilities, to stop publishing public reports.
Why it matters: A government can now switch off the most capable AI models overnight, and the stated reason, a model that cracks classified systems in hours, is exactly what makes that power so consequential. The harder question is who gets to make that call, and whether anyone outside the White House can see how it is used.
TARS DexHand: Giving Robots the Sense of Touch
At ICRA 2026 in Vienna, embodied AI startup TARS is debuting DexHand, a 21-DoF hand with tactile sensors detecting textures down to 0.05mm. Paired with the AWE 3.0 foundation modelโwhich includes a newly released technology called TacForeSight that allows the robot to predict contact events before they fully occurโthe system achieves "hand-brain integration." Live: their A1 robot packs a backpack and performs sub-millimeter wire harness assembly with real-time error correction by using the same framework.
The chart: "API prices of key AI models" (Artificial Analysis, data as of 15 June 2026) plots each model's AA Intelligence Index against its blended price per million tokens, a 7:2:1 mix of cached input, input, and output. Anthropic's Claude Fable 5 tops the index at 64.9 but is the most expensive by far at a $7.7 blended price ($10 in, $50 out). Claude Opus 4.8 (61.4) and GPT-5.5 (60.2) sit near $4, while Chinese labs cluster cheap: MiniMax-M3 scores 54.7 at $0.2, DeepSeek V4 Pro 49.8 at $0.2, and Google's Gemini 3.1 Pro 57.2 at $1.7.
The lesson: The last few points of intelligence cost a fortune. Fable 5 buys only a few points over Opus 4.8 and GPT-5.5 at nearly double their blended price, and runs about 38x the cost of MiniMax-M3 or DeepSeek V4 Pro, models that sit just 10 to 15 points lower. For most workloads, the frontier premium is hard to justify.
The caveat: AA Intelligence is a single aggregate that misses coding, agentic, and domain-specific gaps, and the blended price assumes a 7:2:1 cache ratio few real workloads hit. Heavy-output or low-cache jobs make the priciest models even pricier, and these list prices are not the discounts large customers negotiate.
๐ธ Wall Street's Newest Hire Is an AI Agent
โก Bottom line: JPMorgan now runs its in-house LLM Suite for about 250,000 staff and is scaling AI agents from 450 live use cases toward 1,000 this year.
๐ก Why it matters: The largest US bank is turning AI from a chatbot into software that does real financial work, and rivals are following fast.
๐ What it means: As agents absorb junior tasks, banks are openly rethinking headcount, hiring ratios, and where human judgment still earns its keep.
At the largest US bank by market value, the fastest-growing workforce is software. JPMorgan Chase's in-house LLM Suite, a single interface that routes each task to a suitable model from providers including OpenAI and Anthropic, now reaches about 250,000 employees, roughly two-thirds of staff, onboarded in around eight months. About half use it daily. The bank says it is moving from chat to agentic AI, software that plans and carries out multi-step workflows, and is scaling from 450+ live use cases toward 1,000 this year.
JPMorgan is rolling agentic AI across consumer banking, markets, and wealth management (The Digital Banker)
The numbers are concrete. Generating an investment-banking pitch deck, once a multi-hour grind, now takes about 30 seconds, and employees report saving three to six hours a week. JPMorgan pegs the value at $1.5 billion a year, with AI fraud systems adding roughly $250 million in annual savings and over $1 billion in loss prevention, on an $18 billion technology budget. McKinsey estimates agentic AI could add $200 billion to $340 billion a year across global banking.
Banks are recasting AI agents as core infrastructure, not experiments (Emerj)
What sets this apart from past automation is that the work is judgment-shaped: drafting credit memos, comparing filings, summarizing earnings calls, the bread and butter of analysts and associates. Leadership has openly discussed shifting junior-to-senior banker ratios and trimming operations roles over the next few years. The likelier near-term result is not mass layoffs but a quieter reshaping: fewer entry rungs, more oversight of machine output, and a premium on the relationships, accountability, and proprietary data that, as Satya Nadella argued this month, no model can hand you.