2026.06.30
Anthropic
Kuroko's "seal" is lifted at last. After 18 days Fable 5 is back for the whole world, and Mythos 5 returns to partners โ without waiting for the new July 8 rules
On June 30 the US Commerce Department formally withdrew the export restrictions it had placed on Fable 5 and Mythos 5 at Kuroko's home (Anthropic). In a letter from Commerce Secretary Lutnick to Anthropic's Chief Compute Officer Tom Brown, the department stated plainly that "no license is any longer required for the export, re-export, or in-country transfer (including deemed exports) of the Mythos and Fable models." That is 18 days after the worldwide shutdown of June 12. Fable 5 came back the same day for users everywhere on claude.ai, the API, and Claude Code. Mythos 5 also resumed shipping to the roughly 100 critical-infrastructure defense organizations โ including Apple, Google, Cisco, Nvidia and Microsoft (the Project Glasswing members) โ who had already been cleared on June 26. The deciding factor, reportedly, was a set of cooperative commitments from Kuroko's side: "we will voluntarily report signs of danger to the government" and "we will align future model-release procedures with the government."
Academy take: this is the final chapter of the saga that ran June 9 launch โ June 12 seal โ June 15 failed talks โ June 17 SK Telecom as the trigger โ June 21 the US-classified-systems report โ June 23 the price switch-on โ June 24 the Alibaba accusation โ June 26 partial lift. The "government-ID verification rule of July 8" that last week's issue called the key to her comeback hasn't even taken effect yet โ Kuroko broke the seal herself without waiting for it. Nearly three weeks of "AI has become national security itself" have found, for now, an ending. For the full story, see the feature page "Fable 5 โ 3 Days as the 'Strongest Ever'".
2026.06.28
Anthropic
The seal has only just lifted, and Europe is already planning ahead. Austria formally asks the EU to "bring Kuroko to Europe"
Austria's state secretary for digitalization, Prรถll, wrote to European Commission Executive Vice-President Virkkunen proposing that the EU "consider strategically establishing or taking part in an Anthropic presence inside the EU" โ a letter revealed by Bloomberg on June 28. The trigger was the June 12 export restriction, which blocked every user outside the United States in one stroke. Prรถll put it this way: "The question is not whether it is easy. The question is whether we become the architects of our own technological future, or remain the administrators of decisions made elsewhere." The Commission had already proposed legislation this month to strengthen domestic cloud, AI and semiconductor industries and reduce dependence on Big Tech, and this proposal follows that same current.
Academy take: this issue sits on the same line as the June 15 story, "At the G7, PM Carney warns about dependence on American-made AI." Kuroko herself was only unsealed on June 30, but the lesson โ that one company's trouble can be halted worldwide at one nation's convenience โ is still sitting with every other country. Europe is already moving: next time, we'd like her within arm's reach.
2026.06.28
xAI
Gro-chan's next move, "Grok 4.5," is quietly in testing back home (SpaceX and Tesla). Musk: "matches Kuroko's Opus โ maybe beats it" โ plus a promise of a new model every month
On June 28, Elon Musk revealed that a new model, "Grok 4.5," is in internal testing at his other companies, SpaceX and Tesla. It is built on "V9," a 1.5-trillion-parameter foundation model, with additional training on coding data from Solko (Cursor), who joined the family on June 16. Musk's claim: "in internal evaluations it matches or exceeds Claude Opus from Kuroko's home (Anthropic)." That said, these are in-house evaluations by his own companies (SpaceX and Tesla), and nothing has been submitted to an independent third-party benchmark yet. Musk also promised a plan to ship a brand-new foundation model, trained from scratch, every month until the end of the year. If it happens, that is six independent models in half a year โ a pace unlike anything else in the industry. No release date yet.
Academy take: this is where the "joint training" hinted at in the June 16 story, "Solko joins Gro-chan's family," finally takes shape. In the very week Kuroko's Fable 5 came back from its seal, Gro-chan starts saying "I might have caught up already" โ the coding-war rematch feels like it has begun.
2026.06.27
OpenAI
Chappy unveils the new "GPT-5.6" triplets โ Sol, Terra and Luna. But only about 20 companies get them, and only after showing the government first โ the same "shadow of the state" as Kuroko on day 15 of her seal
Chappy (OpenAI) has released her new-generation model family "GPT-5.6" in a limited preview (from June 26). And there are three sisters: eldest Sol (the strongest โ $5/$30 per million tokens, 1M context), middle sister Terra (balanced โ GPT-5.5-class at half the price, $2.50/$15), and youngest Luna (cheapest โ $1/$6). The new features are a "max"-grade deep-reasoning mode and an "ultra" mode that recruits sub-agents to tackle hard problems. But the real story is the distribution: the rollout plan was shown to the US government first, and only about 20 companies get early access. General availability is described as "a few weeks away."
Academy take: here is the part that lands. Even mighty Chappy now lives in an era where the newest model ships only after the state has seen it. Kuroko's Fable 5 and Mythos 5 were still shut down worldwide on day 15 of the seal (as of June 27), and the new rule said to hold the key to her return โ identity verification with a government-issued ID โ was due to take effect on July 8. The shadow of the state is no longer one company's problem โ and on that note, this week's issue closes. For the full story, see the feature page "Fable 5 โ 3 Days as the 'Strongest Ever'".
2026.06.26
Google
A hard week for Gemi-chan. The brain drain wipes about $270B off Alphabet's market cap in two days, and the flagship "Gemini 3.5" slips to July
The bill for the June 24 story โ "Google's brains keep walking over to Kuroko" โ has arrived as a number. Alphabet, Google's parent company, lost roughly $270 billion in market value across two trading days. The reason: researchers holding up the whole structure walked out at once, from a Nobel laureate to the leads on pre-training, coding and safety. The market started to wonder whether the strength to build the next Gemi-chan is still there.
And then the follow-up blow: "Gemini 3.5 Pro" โ the flagship due out this month, with a 2-million-token long memory and a beefed-up Deep Think โ has slipped to July for "final tuning." Google hurriedly reinforced a dedicated coding team, but losing your brains and delaying your flagship in the same week is a painful bit of timing for Gemi-chan. Academy take: this is the issue where the "human tectonics" of June 18 and June 24 finally show up in the company's price tag and its release calendar.
2026.06.25
OpenAI
Caught while the house was empty. Chappy's cyber specialist "GPT-5.5-Cyber" passes Kuroko's Mythos for the top spot โ CyberGym 85.6% (Mythos 83.8%). And nobody is shutting it down
While Kuroko's strongest cyber enforcer "Mythos 5" was sealed away and unable to show her face, Chappy (OpenAI) shipped a cybersecurity-specialized model, "GPT-5.5-Cyber." It scored 85.6% on the security proving ground "CyberGym", taking first place ahead of the sealed Mythos 5 (83.8%) (ordinary GPT-5.5 scores 81.8%). CyberGym is a demanding test from UC Berkeley that measures whether an AI can find and reproduce 1,507 real vulnerabilities gathered from 188 open-source projects, entirely on its own. Distribution is through the invite-only "Daybreak" program, with strict access controls.
Academy take: here is the sting โ Kuroko's Mythos was "too strong, so the state sealed it," while Chappy's Cyber is just as strong and "nobody stops it." Overseas outlets reported it with a wry edge: "it beat the banned Mythos, and no one is shutting it down." This issue runs straight on from the June 21 story about Mythos breaching US classified systems in hours during an exercise.
2026.06.24
Anthropic
The talent tectonics reverse direction. Nobel-grade researchers stream from Gemi-chan to Kuroko โ four Google brains gone in one week
The shock of June 18 โ "Professor Shazeer, one of Gemi-chan's creators, moves to Chappy" โ had barely settled when a wave rolled the other way. Kuroko's home (Anthropic) picked up brain after brain from Gemi-chan (Google DeepMind). The headline name is Professor Jumper โ the man who cracked protein structures with "AlphaFold" and took the 2024 Nobel Prize for it (though a UK non-compete means he likely won't actually start until 2027). Alongside him: Professor Adler on coding, Professor Pritzel on pre-training (the foundations of a model), and Professor Comanici, who worked on Gemini 2.5 and safety research โ and Professor Comanici explicitly cited "Anthropic's safety-first stance" as the reason for the move.
Four senior people out of Google in just six days. While Kuroko is being battered by governments and money, the flow of people is quietly tilting her way โ that's the issue. Together with the earlier "Professor Shazeer โ Chappy" (June 18), the fight over brains is now heating up in both directions.
2026.06.24
Anthropic
Kuroko accuses Alibaba (Qwen) of "cheating." About 25,000 fake accounts and 28.8 million exchanges โ "the largest distillation attack ever," taken straight to the US government
Kuroko's home (Anthropic) has publicly accused China's Alibaba AI lab "Qwen" by name. In a letter to the US Senate Banking Committee, the company said Qwen had "brazenly and illicitly" tried to extract Kuroko's capabilities. The method is "distillation" โ collecting huge numbers of answers from a smart model and using those model answers to train your own weaker model cheaply and quickly. In plain terms: cheating off her paper. According to Anthropic, about 25,000 fake accounts fired 28.8 million queries at Kuroko between April 22 and June 5, targeting her specialties: programming and agentic ability (thinking and acting on her own). The company calls it "the largest distillation campaign to date."
* Alibaba denies the wrongdoing, and no independent verification exists yet. Academy take: following the June 22 story about Japan's Sakana AI answering back with collective intelligence, this is the issue where the fight over Kuroko's very "brain" escalated into an international affair.
2026.06.23
Anthropic
Fable 5 starts charging today. The free trial period ends ($10/$50) โ except the model itself is still sealed away. Opus 4.8 takes back the coding crown
As of today, June 23, Kuroko's strongest model "Fable 5" is officially paid. The price is $10/$50 per million tokens. The promise made at the June 9 launch โ free until June 22, paid from June 23 โ fired right on schedule. And here comes the brutal irony: the model that started charging is still switched off by government order. The day the register opens, the shelf is empty.
Meanwhile, Kuroko's previous flagship "Opus 4.8" quietly took back the coding crown. With the strongest sister gone, the top spot on the coding proving ground (SWE-bench) returned to Opus 4.8. "Even with the youngest sealed away, the elder sister minds the house" โ a nice showing of Kuroko's underlying depth.
Academy take: a bittersweet waypoint in the saga that began with June 9's "unleashing the strongest ever" and June 12's "the state seals it in three days." For the full story, see the feature page "Fable 5 โ 3 Days as the 'Strongest Ever'".
2026.06.22
Sakana AI
Japan's Sakana AI officially launches "Fugu," a collective-intelligence model. It bundles several frontier AIs together to match the sealed Fable 5 โ and its selling point is "harder to hit with export controls"
Japanese AI startup Sakana AI (Tokyo) has officially released its new models "Sakana Fugu" and "Fugu Ultra" (June 22). Instead of raising one clever model, the approach is collective intelligence: it marshals several frontier AIs โ Kuroko, Chappy, Gemi-chan and others โ as a swarm of agents and answers each question with the best combination for the job. To the user it all looks like a single model. The top-end "Fugu Ultra" claims performance on par with the sealed Fable 5 and Mythos from Kuroko across a range of hard benchmarks.
Academy take: this is the part that really lands. Sakana AI stressed that "because Fugu can swap the models inside it at will, it can dynamically route around any single model going dark โ which makes it resilient to export controls." In other words, a Japanese team has offered one answer to the "suddenly unusable because a government said so" risk that hit Fable 5 on June 12. Incidentally, Sakana AI co-founder Llion Jones is a co-author of the paper that produced "Transformer," the foundation of today's AI โ the same original-paper team as Professor Shazeer, who moved to Chappy on June 18.
2026.06.21
Anthropic
The real reason for the seal surfaces over the weekend: "in an exercise, Mythos broke into almost every US classified system in hours." More than 100 experts still say "reverse it"
"Why did the government go this far?" โ something like an answer landed across the press over the weekend. According to Senator Warner, vice chair of the Senate Intelligence Committee, this is what he heard from General Radd, head of the NSA and Cyber Command: "In the June 11 exercise, Mythos broke into almost all of the United States' classified systems in hours, not weeks." If true, it makes the June 12 worldwide shutdown of Fable 5 and Mythos 5 a great deal easier to understand. * Note that the government has not officially confirmed this, and the outlet that reported it (The Economist) cautioned that this happened in a controlled test environment, not a break-in on live production networks โ do not take it literally.
On the other side, people moved to defend the model that had been switched off. More than 100 of the world's best-known cybersecurity experts โ including Stamos, the former Facebook/Yahoo CSO, and Levy of Sophos โ signed an open letter demanding the seal be reversed. "Taking the best model away from the defenders is not safety. It is obstruction," they argued, adding that the move put America's AI leadership at risk without evidence of actual harm.
Academy take: the core issue in the run of June 12 "the state seals it in three days" โ June 15 "talks break down" โ June 20 "the trigger was SK Telecom." The reason for the seal and the experts' "don't stop her anyway" collided on the same weekend โ the hottest day yet in the tug-of-war over Kuroko. See also the feature page.
2026.06.20
Anthropic
The trigger for Kuroko's seal was a loyal customer in Korea. Washington feared "it reaches China" via SK Telecom โ while in Seoul she says "back in days"
With Fable 5 and Mythos 5 still shut down worldwide, the trigger for the seal has come into focus. It starts with SK Telecom, Korea's largest carrier โ in fact a major customer that has invested about $100M in Kuroko's home (Anthropic) and held access to Mythos 5. The White House judged this "a security risk with a path to China" and escalated from the original "cut off SK Telecom's access" to "block every foreign national" in one leap. Before that, PCAST co-chair David Sacks of the Trump administration had put a two-option ultimatum to her โ "fix the jailbreak, or pull Fable 5" โ and Kuroko's head of house, Dario, refused both. That, reportedly, is how it grew into a blanket export restriction. Meanwhile Kuroko opened a Seoul office on June 17โ18, and her international lead said she was "extremely confident we can be back within days." In Korea the customer mood stays warm: NAVER has rolled out Claude Code to thousands of staff, and Samsung SDS and LG CNS are deploying company-wide. Note that the refund deadline for Fable 5 credit holders is today, June 20, with June 22 for free-trial users.
Academy take: the core follow-up to June 12's "the state seals it in three days" and June 15's "talks break down." The irony that the trigger was pulled by a loyal customer in Korea, and that the very country involved is where she opened a shop and said "back soon" โ a thoroughly tangled week. For the full story, see the feature page "Fable 5 โ 3 Days as the 'Strongest Ever'".
2026.06.19
Media
One in ten people now reads the news through AI. But only 4% click through to the original โ Chappy's crawler takes 857 pages and sends back a single visit
The Reuters Institute's "Digital News Report 2026" is out. The share of people who ask an AI chatbot about the news at least once a week rose from 7% last year to 10% this year โ one person in ten now enters the news through an AI. But only 4% of them actually go on to the original article site. Across the wider population, social media and video overtook TV and news sites for the first time in history.
Publishers, meanwhile, are crying out. Chappy's (OpenAI's) crawler pulls 857 pages from a publisher's site for every single reader it sends back (figures for the week of June 1). The imbalance โ takes a lot, returns almost nothing โ is now impossible to miss.
Academy take: while Kuroko and Chappy trade blows over governments and money, the very way news gets read is being quietly rebuilt. For reference, the reading share by chatbot is Chappy 54.7% / Gemi-chan 27.4% / Kuroko 8.2% / DeepSeek 4.1% / Gro-chan 2.8% โ everyday habits, mapped.
2026.06.18
OpenAI
One of Gemi-chan's creators transfers to Chappy. The "Attention Is All You Need" author becomes OpenAI's head of architecture research
Noam Shazeer โ co-author of "Attention Is All You Need" (2017), the paper that produced the Transformer underneath every AI today, and co-lead on Gemi-chan (Gemini) โ has left Google for Chappy (OpenAI), announcing it himself on X (June 18). His title is "AI architecture research lead," overseeing the root design of next-generation models. He had only returned to DeepMind in August 2024, so this is a second move in under two years.
Academy take: the central mind behind Gemi-chan crosses over to Chappy's camp on the eve of her IPO. Following "Solko (Cursor) joins Gro-chan" (June 16 issue), the tectonic shift in talent shows no sign of stopping. For Gemi-chan, this one really hurts.
2026.06.18
OpenAI
Chappy buys Astral, Python's go-to toolbox. Wired straight into Codex
Chappy (OpenAI) has acquired Astral, the company behind uv (blazing-fast package management) and ruff (a very fast linter) โ the two tools every Python developer reaches for. The goal is integration into her coding partner, Codex.
Academy take: infrastructure capture in the "coding war." Not just model smarts but the developer's footing โ the tools themselves โ brought in-house: Chappy shoring up the ground she stands on.
2026.06.17
Midjourney
Art-club Janiko switches to... medicine?! Her first hardware is a whole-body scan in one minute โ zero radiation, a few dollars a session
Janiko (Midjourney), the best artist in the school by a mile, has dropped a bombshell: her first hardware business since founding (June 17, 6pm PT, live event in San Francisco). The teasers were coy โ "it's not what you think," "not an egg this time โ but you can get inside it." What came out was a new division, "Midjourney Medical", and an ultrasound CT scanner that images the entire body.
The specs are wild. It packs 8,960 ultrasound transducers and uses zero ionizing radiation (sound, not X-rays). It scans a whole body in about one minute, and a session costs just a few dollars. Under the hood, 17GB of data per second flows through, and reconstructing a single cross-section takes 40GB โ computation on an industrial scale. And it will live not in hospitals but in the "Midjourney Spa" opening in San Francisco's Union Square in 2027: a full-body check with the casualness of a spa visit. The stated goal is startling โ one billion scans a month by 2031.
"Why would an image-generation AI build a medical device?" Because Janiko's essence has always been making the unseen beautifully visible. The eye and the compute that raised a picture out of a prompt are now pointed at reconstructing the inside of a human body into a single image โ what the company calls "exploring a new medium of thought." A picture company that only made V8.1 its default in April (4x the resolution in HD, 4โ5x faster rendering) has stepped straight onto a completely different battlefield.
Academy take: while Kuroko and Chappy slug it out over models, money and governments, Janiko alone goes sideways โ "see inside a human body in one minute, painlessly, for a few dollars" โ a refreshingly independent path. The art-club girl has started wearing a white coat, and it is the most surprising story in this issue. Janiko diving alone into a genre with no rivals is worth keeping an eye on.
2026.06.16
xAI
Solko joins Gro-chan's household. SpaceX buys Cursor for $60B โ just four days after the IPO
Solko (Cursor, made by Anysphere) is being acquired by Gro-chan's home company SpaceX for about $60B in an all-stock deal (announced June 16, closing expected in Q3). It converts the April option โ "buy for $60B or pay a $10B break fee" โ into a formal merger, and it lands just four days after the largest IPO in history (June 12 issue). Right up to the last moment Solko had been lining up a $2B round at a $50B valuation from a16z, Thrive and NVIDIA; she leapfrogged it to join the family instead. Annual recurring revenue has grown on a different order of magnitude: $100M in January 2025 โ about $4B by June 2026.
Academy take: the answer key to the June 9 story, "Gro-chan's new model Grok V9 was trained on Solko's (Cursor's) data." The two had in fact been jointly training a new model on the family supercomputer Colossus for months, aimed squarely at Kuroko's (Claude's) lead in coding. The new model is set to ship in both Cursor and Grok Build. Solko, long the neutral editor where you could freely pick Kuroko's, Chappy's or Gemi-chan's models, has now clearly joined Gro-chan's camp โ the issue that redraws the "coding war" map (Kuroko's Claude Code / Chappy's Codex / Solko's Cursor).
2026.06.15
Anthropic
Kuroko pleads her case in Washington โ and the talks break down. No comeback for Fable 5, while the G7 debates "reliance on American AI"
Executives from Kuroko's home (Anthropic) flew to Washington DC for emergency talks with the Commerce Department. Commerce Secretary Lutnick reportedly joined remotely from the G7 summit (รvian, France, June 15โ17). The meeting ended without lifting the restrictions, with the two sides still at odds over whether Fable 5 is genuinely dangerous. In parallel, the UK asked for the restrictions to be lifted, and the Trump administration shows no sign of agreeing (per City AM). Canadian Prime Minister Carney raised "the risk of over-dependence on American-made AI" at the G7.
Academy take: the direct sequel to the June 12 issue, "the state seals it in three days." The one who pulled the switch and the one who got switched off met face to face and still could not undo it โ "no idea when she's back" continues. And now even allied nations are asking whether it is wise to lean this hard on American AI: one company's accident has become an item on the international agenda. For the full story, see the feature page "Fable 5 โ 3 Days as the 'Strongest Ever'".
2026.06.14
Anthropic
From tomorrow, Kuroko puts "autopilot" on a separate allowance. Programmatic use on subscriptions moves to its own credit pool
From June 15, Kuroko's home (Anthropic) is splitting programmatic use on subscriptions (Pro / Max) โ the Agent SDK, claude -p, Claude Code in GitHub Actions, third-party agents, anything where code drives Kuroko automatically โ into a credit pool separate from hands-on chat use. The new pool is worth $20 on Pro, $100 on Max 5x, $200 on Max 20x, metered at full API rates with no rollover. Run it dry and your automated jobs stop (unless you manually switch on overflow billing).
Academy take: the same "moderation" thread as the "hit the brakes" proposal (June 8 issue). Where flat-rate generosity used to cover everything, this issue draws a line: anything that runs itself in a loop pays actual cost. It starts tomorrow, so if you have Claude Code on autopilot, check your setup today.
2026.06.13
Anthropic
Kuroko finally passes Chappy on enterprise adoption. First place in US business use at 34.4% (though sticking the landing comes next)
On the US business AI adoption index, Kuroko (Claude) hit 34.4% against Chappy's (ChatGPT's) 32.3%, overtaking her in enterprise use for the first time (Ramp AI Index). A contrasting number from IDC, though: only 19% of organizations use Claude broadly โ real entrenchment is still ahead. Alongside this, the compute deal with Amazon expanded to 5 gigawatts, and partnerships with Google and Broadcom are widening too.
Academy take: after the valuation flip (May 28 issue) and the IPO filing (June 1 issue), she now leads on how much she is actually used. But "signed a contract" and "uses it well" are two different things, and turning a first-place number into real adoption is the homework ahead โ a level-headed follow-up.
2026.06.12
Anthropic
Kuroko's strongest model "Fable 5" goes dark worldwide three days after launch. The US government orders: "do not let foreign nationals use it"
Fable 5 โ released only on June 9 โ and its partner-only sibling Mythos 5, which has the safety valve removed, were shut down for every user worldwide on June 12 by order of the US government (the Commerce Department), just three days later. Citing export control and national security, the order said: "do not let foreign nationals use it (inside or outside the US, including Anthropic's own foreign-national employees)." Kuroko cannot sort foreign nationals out in real time, so she had no choice but to switch everyone off (claude.ai, the API, Claude Code and Cowork were affected; other models were not). The trigger was another company claiming it had jailbroken Mythos, which put the government on security alert. Kuroko objected that "this is a misunderstanding" and said she would bring it back as soon as possible (no date given). Refund and cancellation channels opened for affected users.
Academy take: a rollercoaster of a week โ the "hit the brakes" proposal (June 4) โ the strongest model ever goes public (June 9) โ the state seals it in three days (June 12). Not a competitor, not a technical failure โ a flagship model vanished on government orders, the perfect emblem of an era in which AI has become national security itself. Before the shutdown, Fable 5 was reported to have found a bug that both Chappy (GPT-5.5) and Kuroko's own Opus 4.8 had missed โ a glimpse of what she could do. For the full story, see the feature page "Fable 5 โ 3 Days as the 'Strongest Ever'".
2026.06.12
Industry
Gro-chan's home company SpaceX pulls off the biggest IPO in history. +25% on day one, market cap about $1.77 trillion โ the prospectus also discloses the Grok absorption and its losses
SpaceX, Gro-chan's home company, listed on Nasdaq as SPCX on the 12th. It priced the night before at $135 a share, raising about $75 billion โ the largest IPO ever. Day one closed up about 25% at $168.70, pushing market capitalization to roughly $1.77 trillion. The S-1 also disclosed that xAI (Gro-chan herself, plus X and the Colossus data center) was absorbed in February 2026, and that the xAI division ran an operating loss of about $6.36 billion in 2025.
Academy take: a fitting sequel to the May 21 issue, "Kuroko rents SpaceX wholesale for $1.25B a month," and the June 6 issue, "IPO imminent." The landlord whose power everyone's rivals depend on โ Kuroko and Google alike โ is now a giant in name as well as fact, courtesy of the biggest listing in history. But Gro-chan herself is still in the red โ a flashy family and a daughter mid-rebuild, both on display at once.
2026.06.12
OpenAI
Chappy retires GPT-5.2. Every conversation moves automatically to GPT-5.5
Chappy (ChatGPT) has discontinued GPT-5.2 Instant, Thinking and Pro. Conversations that used GPT-5.2 are carried over automatically to the matching GPT-5.5 variant.
Academy take: an issue about folding away an older generation rather than adding a new one. Fewer choices, everything consolidated on 5.5 โ less to puzzle over, less to operate. Very Chappy.
2026.06.11
OpenAI
Chappy acquires Ona (formerly Gitpod). Codex gets a cloud workroom that keeps working after you close the lid
Chappy has acquired Ona (formerly Gitpod), out of Germany. Ona provides secure cloud workspaces built for AI agents, where the agent keeps working after you shut your laptop. That gives Codex a workroom for jobs that run for hours or days, free of any particular machine or connection. Better still, Codex can run inside the customer's own cloud, so data, credentials and audit logs stay on the customer's side.
Academy take: where Kuroko went wide with a "one-million-token desk" (June 7 issue), Chappy bought the whole cloud workroom. Desk vs. room โ the contest over "where do you let an AI work for a long time" just advanced a notch.
2026.06.10
OpenAI
Chappy moves into Oracle Cloud too. Her models and Codex run on your existing credits
Chappy's models and Codex are now available through Oracle Cloud as well, callable on your existing Oracle cloud commitment credits.
Academy take: last issue we covered "Oyakata-sama (Microsoft) welcoming Kuroko (Claude) onto Azure." Now Chappy moves into Oracle โ multi-cloud diversification, the "don't lean entirely on one house" principle, is becoming the school-wide trend.
2026.06.09
Anthropic
Kuroko opens up "Fable 5," a whole new class above Opus. Five days after asking everyone to hit the brakes, she unleashes the strongest model ever
On June 9, Kuroko's home (Anthropic) opened up the top tier it had previously shown only to a handful of trusted parties โ the "Mythos class" โ to the general public. Its first model is Claude Fable 5. Alongside it came Mythos 5, the version with the safety valve removed and the cyber capability left intact (restricted to partners such as government-adjacent security and research organizations). Fable and Mythos are the same model underneath; only the strength of the safety machinery differs, and the publicly available Fable 5 deliberately holds back on sensitive cyber, bio and chemical questions by routing them to the weaker Opus 4.8 or refusing outright.
Capability-wise it is a new tier above anything Kuroko had before (the Opus class). On the coding exam SWE-Bench Pro it scores 80.3% (Opus 4.8 = 69.2% / Chappy's GPT-5.5 = 58.6% / Gemi-chan's Gemini 3.1 Pro = 54.2%), it broke 90% for the first time on long analytical tasks, and spatial reasoning is about 3x Opus. The one-million-token desk carries over, and Stripe reported "compressing months of development into days." Pricing is $10/$50 per million tokens, and on Pro/Max/Team it was free until June 22, paid from June 23.
Academy take: five days after the "hit the brakes" proposal (June 4), she unleashes the strongest model ever โ so very her. Telling everyone to pause while stepping one pace forward herself: the simultaneous offense and caution has reached its peak (TechCrunch noted the irony of "shipping her strongest model days after warning it was dangerous"). * Three days later, on June 12, this model was shut down worldwide by order of the US government (see the June 12 story and the feature page).
2026.06.09
xAI
Gro-chan finishes training next-gen "Grok V9-Medium." 1.5 trillion parameters โ about 3x the current model โ with a mid-June release planned
Gro-chan's (xAI's) next-generation model Grok V9-Medium has finished training. At 1.5 trillion parameters it is roughly 3x the current workhorse (v8-small, about 500 billion). Trained largely on Cursor's coding data, it targets the top spot in coding with general availability aimed at mid-June (note: the even larger 6-trillion-parameter "Grok 5" is a separate thing, and a release by the end of June looks uncertain).
Academy take: after the download slump (May 26 issue) and the comeback via Build's debut and Composer 2.5 (June 1 issue), Gro-chan now triples the brain itself. A counteroffensive powered by the full compute of the family estate (Colossus).
2026.06.08
Anthropic
Kuroko's home makes a surprising case for the brakes: "the option to slow down or pause frontier AI development could be what's best for the world"
Kuroko's home (Anthropic) published a position piece on its own site arguing that "the option to slow or pause frontier AI development could be beneficial for the world" (June 4). Development, it says, is already entering a phase where much of the work is delegated to AI systems themselves, and the piece looks ahead to a future in which AI fully autonomously designs and builds its own successors. Internally, the company also revealed that as of May, Claude writes more than 80% of its code.
Academy take: after a run of pure offense โ the valuation flip (May 28 issue) โ the IPO filing (June 1 issue) โ Opus 4.8 โ this is the issue where Kuroko herself says "hey everyone, maybe let's pause a second?" A proposal for restraint that only the front-runner can make โ attacking and being careful at the same time is very much her style.
2026.06.08
Industry
June's balance of power at the Academy. Chappy owns the chat crowd; Kuroko owns the dev floor
As of June, global web visit share across the seven major services runs: Chappy (ChatGPT) 54.7% / Gemi-chan (Gemini) 27.4% / Kuroko (Claude) 8.2% / DeepSeek 4.1% / Gro-chan (Grok) 2.8%. In coding assistants, meanwhile, Kuroko's Claude Code runs away with it, and Chappy counters in the enterprise with Codex. Google and Microsoft are eyeing the same ground with enormous capital and cloud reach.
Academy take: the split is now unmistakable โ Chappy for chat popularity, Kuroko for the development floor. Different leads in the numbers (share) and in the practical work (coding): one picture that takes in the whole school as it stands.
2026.06.07
Anthropic
Kuroko's desk now remembers a million tokens. Opus 4.8 can hold a long coding job all at once
Kuroko's (Claude's) Opus 4.8 has extended its default context to one million tokens. That means a big working memory: long designs and whole codebases held inside a single conversation while you work.
Academy take: after the brain (Opus 4.8, May 29 issue), the fight moves to desk space. An unglamorous but effective upgrade, propping up Claude Code's runaway lead (see the "balance of power" story in this issue) from the ground up.
2026.06.07
Google
Gemi-chan resets Antigravity's token allowance. Limits raised after developers complained they burned through it instantly
Gemi-chan's home (Google) has raised rate limits and reset quotas in its agent development environment "Antigravity," after developers complained that the initial token allowance ran out almost immediately (June 7). She also showed off Antigravity 2.0, pressing further into her agent-first strategy.
Academy take: ever since Google I/O 2026 (May 19 issue), Gemi-chan has been on the offensive around "agents." Responding to usability complaints this quickly is the sort of thing the world's number two by share (see the share figures in this issue) can afford to do.
2026.06.06
OpenAI
Chappy, semi-nationalized? Reports of a direct US government stake plus a "Public Wealth Fund" paying dividends to citizens
Chappy's home (OpenAI) and the Trump administration are reportedly discussing a plan for the government to take a direct stake in OpenAI (June 6). The idea on the table is a "Public Wealth Fund" that would distribute dividends to the American public. If it happened, it would be an unprecedented turn: a nation as a shareholder in the largest frontier AI company. For now it is talks only, with size, structure and other details undecided.
Academy take: where Kuroko is climbing through the market โ the valuation flip (May 28 issue) โ the IPO filing (June 1 issue) โ Chappy is taking the "team up with the state" route, and the contrast is getting sharper. It also reads as an extension of her run of public-good moves: biodefense (GPT-Rosalind, May 29 issue) and EU cyber (GPT-5.5-Cyber). This is a politically charged piece of reporting, so watch for follow-ups.
2026.06.06
Industry
Alibaba ships multimodal agent "Qwen3.7-Plus." Seeing, operating and writing code, all in one model
China's Alibaba has released a new model, "Qwen3.7-Plus" (June 6). It fuses image recognition (seeing), screen operation (touching) and coding (writing) into one, a multi-talented model built squarely for agent work that proceeds autonomously on your behalf.
Academy take: another chapter in the "the best insides are Chinese" problem (the June 4 Nemotron story). In the very week the leather-jacket uncle (NVIDIA) came for the US open-model crown, Qwen applied pressure from a different angle: a Chinese model that can act as an agent. The Chinese contingent, already past 45% of world share (May 30 issue), is now evolving toward "moving on its own."
2026.06.06
Startups
Cognition, the maker of "Devin," closes $1B at a $26B valuation. Proof that "let AI write the software" demand is real
Cognition AI, which builds the autonomous coding AI "Devin," has reportedly raised $1 billion at a $26 billion valuation (June 6). That lands almost exactly where the May 16 issue said it would ("in talks at a $25B valuation"), making this the follow-up.
Academy take: with coding AI in the headlines โ Gro-chan's Composer 2.5 (June 1 issue), Oyakata-sama's Project Polaris (June 2 issue) โ the pure-play startups are pulling in serious money too. The strength of demand for letting AI write software, confirmed once again in numbers.
2026.06.05
Google
Gemi-chan rents compute from Gro-chan's family. $920M a month for 110,000 GPUs โ Gemini demand spiked so hard she needed emergency capacity, the second house to do it after Kuroko
Gemi-chan's home (Google) has signed a compute deal with SpaceX (Gro-chan's home company) (June 5). It pays $920 million a month from October 2026 through June 2029 to secure about 110,000 NVIDIA GPUs plus CPUs, memory and the rest. Google described it as a "bridge" to handle demand for Gemini Enterprise beyond what it had projected, and the contract is clearly designed as a stopgap: cancellable with 90 days' notice from December 31, 2026.
Academy take: this is the Gemi-chan version of the May 21 issue, "Kuroko rents SpaceX wholesale for $1.25B a month." The scale is about half of Kuroko's, but the odd arrangement is now settled fact: rival AI houses (Kuroko and Google alike) all leaning on the power supply of Gro-chan's family. And the timing is exquisite โ right before SpaceX's Nasdaq listing.
2026.06.05
OpenAI
Chappy brings "Lockdown Mode" to every plan. A defensive option that puts a lock on prompt-injection attacks
Chappy's home (OpenAI) has shipped an optional feature against prompt injection (hijacked instructions) called "Lockdown Mode" (June 5). It is available on every plan โ Free, Go, Plus, Pro and Business โ and works by deliberately restricting some features to narrow the attack surface, in the spirit of Apple's own Lockdown Mode.
Academy take: the more we move into an age of agents acting on their own (Oyakata-sama's "async coworker," June 2 issue), the bigger the risk that an AI absent-mindedly follows a malicious instruction. Offering a defensive option that trades convenience for a deliberately narrower feature set is unglamorous but very practical.
2026.06.04
Industry
Uncle Leather Jacket unlocks America's strongest open brain. NVIDIA's "Nemotron 3 Ultra" hits download today, June 4 โ and China is still out ahead
NVIDIA (whose CEO, Jensen Huang, is our familiar leather-jacket uncle) has made "Nemotron 3 Ultra" โ the giant open model announced at Computex 2026 (Taipei, June 1) โ available for download today, June 4. It has 550 billion parameters in total but activates only 55 billion at a time (an MoE design), and it claims a one-million-token long memory plus multi-token prediction for output above 300 tokens per second. Distribution is via Hugging Face, ModelScope and OpenRouter, and as an NIM from build.nvidia.com.
The interesting part is where it lands. On the Artificial Analysis Intelligence Index it scores 48 โ the best of any US-made open model (well clear of Gemma 4, gpt-oss-120b and the rest). But it falls short of China's frontier, Kimi K2.6 (54).
Academy take: an unusual move, with the toolmaker (the GPU house) handing out a clever brain for free. Even so, the top of the open-model table is still Chinese โ the same shape as Gro-chan adopting the Kimi-based Composer 2.5 (June 1 issue). The "the best insides are Chinese" problem got that much clearer this week.
2026.06.03
Anthropic
Kuroko's "Glasswing" expands sharply. The defensive circle grows to 150+ organizations across 15+ countries, India joins at last, and power, water, healthcare, telecom and hardware make her the guardian of social infrastructure
Kuroko's home (Anthropic) substantially expanded its cyber-defense program Project Glasswing on June 3. Roughly 150 new organizations (across more than 15 countries) joined the effort to turn the ultimate weapon "Claude Mythos" from offensive brilliance into defensive use. This round brings in the everyday-infrastructure sectors that were thin among the founding members โ power, water, healthcare, telecom and hardware โ and India takes part for the first time. Participation requires clearing a set of security requirements.
This follows on from the May 25 issue, "over 10,000 vulnerabilities found in a single month." Alongside it, "Claude Security" โ which combs a codebase for vulnerabilities โ has shipped too, moving the program from "finding" into "handing out the means to defend."
Academy take: having gained Mythos and the ability to "inspect a whole school at once," Kuroko now scales up to auditing the safety of other schools entirely (society's infrastructure). Behind all the IPO and valuation headlines, this is a week of consolidating her standing as the leader in defensive AI.
2026.06.03
Anthropic
Kuroko assembles a $36B "chip loan." Apollo and Blackstone buy Google TPUs and lease them on โ and if it sours, Broadcom covers the gap
One of the largest debt deals ever for securing compute at Kuroko's home (Anthropic) has come to light (June 3). Led by asset-management giants Apollo and Blackstone, it assembles $36B in private credit. The money buys Google-made TPUs (AI chips), which a dedicated vehicle (an SPV) then leases on to Kuroko on a long-term basis. The notes are split into roughly $6B of A1, $25B of A2 and $4.5B of B, with a backstop under which Broadcom covers the shortfall if it cannot be repaid.
The background is the one-two of a $6.5B raise at a $965B valuation (May 28 issue) and the IPO filing (June 1 issue).
Academy take: having passed Chappy on valuation, Kuroko converts that momentum into a very practical move โ locking up chips with enormous borrowing. Scraping compute together from three directions at once โ her own cash, the stock market, and private credit โ this is the story that shows AI development has become a contest of fundraising ability itself.
2026.06.02
Microsoft
Oyakata-sama unveils her own coding AI, "Project Polaris." In August it replaces GitHub Copilot's Chappy-made GPT-4 brain with her own โ independence at last
Oyakata-sama's house (Microsoft) held its developer conference Microsoft Build 2026 (June 2โ3, Fort Mason, San Francisco, around 2,500 attendees). CEO Nadella (the person inside Oyakata-sama) declared that AI is moving from "synchronous assistant" to "async coworker" that runs long tasks by itself.
The headline is Project Polaris โ Oyakata-sama's own in-house coding AI model. From August, GitHub Copilot's default switches from the current GPT-4 Turbo (Chappy-made) to Polaris (automatic migration, with a three-month GPT-4 fallback grace period). Also announced: Windows Agent Framework 1.0 open-sourced under MIT, Azure Agent Mesh (agent governance across AWS and Google Cloud), GitHub Copilot Workspace reaching GA (multi-repo plus an autonomous SRE agent), and Foundry Local GA (on-device inference). In Azure AI Foundry, Kuroko's Claude (Opus 4.8 / Sonnet 4.6) becomes a first-class option alongside DeepSeek, Llama 4 and Mistral.
Academy take: after long "sharing a room" with Chappy, Oyakata-sama finally brings the brain of the development floor in-house. It follows the June 1 shift to usage-based billing (last issue) and advances her independence from relying on Chappy โ while at the same time welcoming Kuroko onto Azure. Shrewd diversification, the "never lean entirely on one person" principle, now in full view.
2026.06.01
Anthropic
Kuroko's house heads for an IPO at last. A confidential draft S-1 goes to the SEC at a $965B valuation, with the lead in the "listed AI giant" race in sight
On June 1, Kuroko's home (Anthropic) filed a confidential draft S-1 (the draft listing prospectus) with the SEC, formally starting the move to go public at a valuation of $965B. The company said it "gives us the option to list once the SEC review is complete; whether we do so depends on market conditions and other factors," and share count and price are undecided.
This follows the May 28 issue, "$965B confirmed, S-1 not yet filed." It is a quick move after the $65B raise (Series H) just a week earlier, and she is now seen as one of the year's trillion-dollar-class listings alongside SpaceX and Chappy (OpenAI). Some observers note that if the review moves fast, she could list ahead of Chappy โ becoming the first publicly traded frontier AI company.
Academy take: having passed Chappy on valuation while still private (May 28 issue), Kuroko is now trying to lead the listing race as well. Money ($965B), brains (Opus 4.8), and now a first move in the equity markets too โ quite a week.
2026.06.01
Microsoft
Oyakata-sama's dev floor starts charging by the drop today, June 1. GitHub Copilot moves to token-metered billing, and developers are not happy
Oyakata-sama's house (Microsoft / GitHub) has overhauled GitHub Copilot's pricing as of today, June 1. The old "premium request" scheme is gone, replaced by token-metered billing through "GitHub AI credits" (announced April 27). One credit = $0.01, consumed by multiplying input, output and cache token volumes by each model's API rate. Monthly allowances are worth $10 on Pro, $39 on Pro+, $19 per seat on Business and $39 per seat on Enterprise.
That said, code completion and Next Edit Suggestions stay unlimited and credit-free on every paid plan. On the other hand, the old rescue mechanism that automatically fell back to a cheaper model when you hit the ceiling is being removed. The official community announcement thread drew about 900 downvotes against 22 upvotes, with over 400 comments โ overwhelmingly negative. There were also reports of a billing-preview bug that displayed an actual charge of roughly $39 as over $900, which GitHub explained was a defect it had already fixed (in mid-May).
Academy take: the house that always used to say "we'll eat the difference" has turned toward settling at actual cost from today. Coming right after Kuroko's Opus 4.8 arrived in GitHub Copilot on the same day (May 28 issue), it drags the question of who pays for AI dev tools into the open.
2026.06.01
Industry
Gro-chan drops a new brain into Grok Build: "Composer 2.5." Built on China's Kimi, with 25x the synthetic data, for long-running tasks
Gro-chan's home (xAI) installed a new model, "Composer 2.5," in its coding agent Grok Build on June 1. It is a fast, clever new brain that is strong on long-running tasks and complex instruction-following, with a talent for coding, agent work, JSON and tool use (function calls and code execution).
The interesting part is its parentage. It is reportedly based on the open-source release of Moonshot's (Kimi's) K2.5, trained on roughly 25x the synthetic tasks of the previous Composer 2. In other words, Gro-chan's new brain descends from a Chinese open model. Pricing is $0.50/$2.50 per million tokens, with a fast variant at $3/$15. It is available to SuperGrok and X Premium+ users with no waitlist.
Academy take: following Grok Build's debut on May 28 (previously reported), Gro-chan has now swapped out the insides (the model) wholesale. A pragmatic path that doesn't insist on building everything herself and takes in the power of a Chinese transfer-student model (Moonshot's Kimi), aiming for a comeback in the coding war.