Hana AI Academy School Newspaper Back Issues

2026.05.30 OpenAI

Chappy owns up to her "bullet-point overload." A refreshed default model, GPT-5.5 Instant, makes answers tidier — and o3 and GPT-4.5 head for graduation

Chappy's house (OpenAI) has refreshed GPT-5.5 Instant, the default model behind ChatGPT (reported across outlets around May 30). Readability was the headline goal: answers now use roughly 30.2% fewer words and 29.2% fewer lines. The update also improves sycophancy resistance, factuality, and multilingual performance. Research lead Michelle Pokrass described the previous version as "too bullet pilled."
On the flip side, Canvas no longer works with GPT-5.5 Instant or Thinking — prose and code are now handled directly in chat via "writing blocks" and "code blocks." A generational changing of the guard was announced too: GPT-4.5 retires from ChatGPT on June 27 (30 days out) and o3 on August 26 (90 days out), though o3 sticks around in the API.
Academy take: after a week spent playing defense with things like biodefense (see the May 29 story), Chappy turned the inspection inward and checked how she actually feels to use. A rare week of subtraction — trimming back the bullet points she'd piled up.

Sources: Thurrott / The Decoder / OpenAI

2026.05.30 Industry

A leaked Meta memo reveals a three-pillar wearables strategy: an AI pendant, always-recording glasses, and enterprise wearables

An internal Meta memo obtained by US outlet The Information (written by wearables VP Alex Himel) leaked on May 30, laying out a three-pillar strategy for AI devices you wear. The pillars: (1) an AI pendant that records, summarizes and searches your conversations (built on the Limitless technology Meta acquired in 2025, internal testing planned for spring 2027, possibly with a camera), (2) "supersensing" smart glasses with always-on cameras and sensors, and (3) a business subscription called "Wearables for Work."
The memo also sketches a glasses roadmap: "Modelo" in June 2026, "Luna" and "RBM2 Refresh" in the fall, and "Mojito VIP" in December, all running on Meta's "Muse Spark" AI plus an unannounced agent codenamed "Hatch." The sales target is 10 million wearables in the second half of 2026 (up from over 7 million in 2025). Subscriptions are Meta One Plus ($7.99) and Meta One Premium ($19.99). The catch: Reality Labs lost roughly $19 billion in 2025, so the up-front bet remains heavy. ※ All of this is leaked, still-at-the-planning-stage information.
Academy take: Meta has no student enrolled at Hana AI Academy, but the shape of the bet is clear. While everyone else competes over "AI inside an app," Meta is going for "an AI that always watches and listens" — i.e. hardware you wear. The main battlefield for AI may be moving from inside the screen to your collarbone and the bridge of your nose.

Sources: The Decoder / Engadget / The Tech Portal / heise online

2026.05.29 OpenAI

Chappy takes up guard duty for biodefense, opening life-science AI "GPT-Rosalind" to governments and research institutes

Chappy's house (OpenAI) announced a new program called "Rosalind Biodefense" on May 29. It opens up GPT-Rosalind — the life-sciences model released in April, which reasons about molecules, proteins, genes and disease biology far more deeply than a general GPT — to vetted developers plus the US government and allied partners. Intended uses are public-health work: pandemic preparedness, early detection, epidemiological modeling, screening, and non-pharmaceutical interventions (NPIs). Launch partners are Lawrence Livermore National Laboratory, the Johns Hopkins Applied Physics Laboratory, and vaccine coalition CEPI, with the White House and federal agencies briefed in advance. The model is safety-trained to refuse dangerous bioweapons work, and access is deliberately kept narrow so oversight actually bites.
Academy take: after a month of shoring up everyday-life infrastructure (personal finance, Excel, voice), Chappy has swung round to national-security-grade "defensive AI." Kuroko waves the flag on cyber defense (Mythos / Glasswing); Chappy is now staking out biodefense. The model is named after Rosalind Franklin, whose work helped uncover the structure of DNA.

Sources: OpenAI / Axios / The Decoder / Seeking Alpha

2026.05.28 Anthropic

It's official: Kuroko raises $65B at a $965B valuation, formally passing Chappy and closing in on a $1 trillion private company

Kuroko's house (Anthropic) formally closed its $65 billion Series H on May 28 at a post-money valuation of $965 billion. Altimeter, Dragoneer, Greenoaks and Sequoia led the round, and annualized revenue (ARR) hit $47 billion this month — more than triple February's $14 billion in just three months. That formally puts Kuroko past Chappy's house (OpenAI), valued at $852 billion as of March, making Anthropic the most valuable private AI company in the world — and one step away from being the first $1 trillion private company ever. An IPO in an October window is being talked about (no S-1 filed yet).
Academy take: this is the follow-up to the May 26 story about "$900B-plus, $30B in talks" — now confirmed and raised to $965B and $65B. Last week's upset became this week's audited scoreboard.

Sources: Anthropic / TechCrunch / Bloomberg / CNBC

2026.05.28 Anthropic

Kuroko's new model "Claude Opus 4.8" arrives just 41 days after 4.7 — 96.7% on math, a quarter as many missed code defects, and "Dynamic Workflows" running hundreds of agents in parallel

Kuroko's house (Anthropic) released its newest model, Claude Opus 4.8, on May 28 — a mere 41 days after Opus 4.7. Lining up Kuroko's current roster of numbers: agentic coding goes 64.3% → 69.2%, multi-domain reasoning with tool use 54.7% → 57.9%, agentic computer use 82.8% → 83.4%, knowledge work 1753 → 1890, Terminal-Bench 2.1 hits 74.2% (+8.4pt), and USAMO 2026 math jumps 69.3% → 96.7% — the biggest math leap any Opus has made.
On top of that, the odds of it missing a defect in its own code drop to about a quarter of before, and Anthropic is pitching sharper judgment, honesty about progress, and long stretches of autonomous work (some reviewers say it approaches the alignment quality of the "ultimate weapon," Mythos). Pricing is unchanged ($5 / $25 per million tokens), and Fast mode is 2.5x faster at a third of the previous cost. On claude.ai you can now dial the "effort" level. The showpiece is Dynamic Workflows (research preview), which runs hundreds of subagents in parallel within a single session to automate end-to-end code migrations spanning hundreds of thousands of lines. It shipped to GitHub Copilot the same day.
Academy take: Kuroko set her all-time best on funding ($965B) and on brains (Opus 4.8) on the very same day. A picture-perfect peak of a day.

Sources: Anthropic / TechCrunch / VentureBeat / GitHub Changelog

2026.05.28 Industry

Gro-chan enters the coding wars. "Grok Build" hits API beta, powered by her family's famous compute muscle

Gro-chan's house (xAI) launched its coding agent "Grok Build 0.1" in public API beta on May 28 (100+ tokens/sec, $1 per million input / $2 per million output). A CLI version had already gone into early beta on May 25 for SuperGrok and X Premium Plus subscribers ($300/month, or an introductory $99/month for six months). It's built on Grok 4.3, runs up to eight subagents in parallel, and offers 2 million tokens of context in the CLI (256K via API). Selling points include Plan Mode, which shows you the plan before it acts, and a local-first design that never ships your code off the machine. With that, she has formally joined the coding-agent war against Kuroko's Claude Code, Chappy's Codex, and Solko (Cursor).
Academy take: Gro-chan, whose app downloads were reported to be stalling (see the May 26 story), is counter-attacking on developer turf with the weapon she knows best — the raw power of the family compound (Colossus).

Sources: xAI / CIO Dive / eWeek / Engadget

2026.05.28 Industry

CNN sues Papu-chan, alleging 17,000 works were copied without permission. The copyright case lands in a New York federal court

News giant CNN sued Papu-chan's house (Perplexity) in the Southern District of New York on May 28. The complaint alleges that Perplexity copied more than 17,000 CNN works (articles, videos and images) without permission and fed them into the RAG (retrieval plus generation) pipelines behind the Comet browser, Search API, Agent API and answer engine. The heart of the claim: it then serves competing content that is "identical or substantially similar."
Academy take: Papu-chan was riding high on total platform coverage (Comet's arrival on iPhone, reported May 26) — and now catches a copyright headwind from a major news organization. What's really on trial here is how AI gathers its information in the first place.

Sources: CNN Business / ppc.land / U.S. News

2026.05.28 Google

Gemi-chan tweaks her new usage limits on the spot: Flash-Lite goes free and Pro gets a per-prompt cap after complaints

Gemi-chan's house (Google) has already adjusted the new usage limits introduced at I/O 2026 (reported May 19), with fixes landing May 28. Responding to complaints that complex prompts burn through the allowance instantly, Google added (1) a cap on how much quota a single prompt can consume on Pro, (2) free usage for 3.1 Flash-Lite (it no longer counts against your allowance), and (3) a fix for a bug where one or two Omni videos exhausted the quota, plus double the generation allowance for Ultra members. A detailed usage dashboard is planned next.
Academy take: Gemi-chan went on the full offensive at I/O, then, hearing the new rules land badly, turned around and fixed them right away — a nimble week.

Sources: 9to5Google

2026.05.27 Industry

Seek-chan's country restricts the "export" of AI brains: China now requires government approval for private-sector AI talent to travel abroad

Chinese authorities now require government pre-approval for overseas travel by top talent at private AI companies including DeepSeek (Seek-chan) and Alibaba, Bloomberg reported (May 26, picked up widely on May 27). The rule covers founders, researchers and executives, selected individually based on their value to China's AI strategy rather than by job title or employer. In March 2025 the guidance was merely "you may want to avoid trips to the US"; it has now hardened into a formal pre-approval regime that applies regardless of destination. DeepSeek founder Liang Wenfeng already declined an invitation to February's AI summit in Paris.
Academy take: this is the sequel to "the Chinese transfer students now hold 45% of the world" (reported May 26). Having grown up fast, those transfer students are now being fenced in by their own country, which has decided not to let them out.

Sources: Bloomberg / The Japan Times / TechTimes

2026.05.26 Anthropic

Kuroko's house passes Chappy in valuation for the first time, raising $30B at over $900B to become the world's most valuable AI startup

Kuroko's house (Anthropic) is raising $30 billion at a valuation north of $900 billion, and the round is expected to close as soon as this week. If it does, it would top the $852 billion valuation Chappy's house (OpenAI) held as of March for the first time, making Kuroko "the most valuable AI startup in the world."
Behind it is a frankly startling growth rate. Anthropic's Q2 revenue is projected at $10.9 billion, more than double the previous quarter, and the first operating profit in company history is within reach. Sequoia, Dragoneer, Altimeter and Greenoaks are each in for around $2 billion.
Academy take: Kuroko, forever whispered about as "the runner-up," has finally rewritten Chappy's report card this week. After the run of reinforcements — Professor Karpathy's transfer (reported May 20), the SpaceX compute deal (May 21), the Glasswing results (May 25) — this valuation flip reads like the final tally. Chappy is heading out into the public markets with an IPO filing (reported May 22), while Kuroko overtakes her on valuation while staying private: two strikingly different strategies, side by side in one week.

Sources: Bloomberg / CNBC / TechTimes

2026.05.26 Industry

China's transfer-student squad quietly takes 45% of the world. DeepSeek, Kimi and Qwen deliver "the same performance at 1/50th of Chappy's price"

Chinese models (DeepSeek V4, Kimi K2.6, Qwen3.6, GLM-5.1) now account for more than 45% of all tokens processed on OpenRouter, the world's largest model-sharing platform. Just two years ago, in October 2024, their share was a mere 1.2%; it hit 10% in March 2025, 25% in Q3 2025, and now over 45% in May 2026. These are no longer "niche options for special cases" — they are mainstream production workloads around the world.
The standout is DeepSeek V4 Pro. On the composite benchmark "BenchLM" it scores 87 against GPT-5.4's 88 — right on its heels — at roughly 1/50th of GPT-5.4's price. On May 22 DeepSeek made its 75%-off pricing permanent ($0.435 per million input tokens, $0.87 per million output), resetting the industry's price floor. Coding specialist Kimi K2.6 tops the SWE-bench leaderboard among open-weight models, while Qwen3.6 leans on Alibaba's hybrid architecture for efficiency.
Academy take: Chappy, Kuroko and Gemi-chan ran the year as "the big three" — and then a group of Chinese transfer students whose names nobody quite knew ended up occupying nearly half the seats. "Same grades, one-fiftieth the tuition" is a shocking value proposition, and the big three's footing is starting to wobble. This looks like the week the industry's price collapse really began to accelerate.

Sources: Digital Applied (OpenRouter Rankings) / BenchLM / CoderSera

2026.05.25 Anthropic

Kuroko's Mythos Preview finds over 10,000 vulnerabilities in a month. Project Glasswing publishes results spanning Cloudflare, Mozilla and 1,000 open-source projects

Kuroko's house (Anthropic) published a one-month update on Project Glasswing on May 25. Vulnerability hunting with roughly 50 partners using the unreleased Claude Mythos Preview model turned up more than 10,000 high and critical severity vulnerabilities in total. The headline numbers: 2,000 at Cloudflare (400 of them high or critical) and 271 in Firefox at Mozilla (ten times what older Claude models found). Anthropic separately scanned over 1,000 open-source projects on its own and identified 6,202 high and critical vulnerabilities, 1,752 of which have been peer-reviewed by six independent security research firms.
Glasswing turns Mythos's "offensive cleverness" around and points it at defense — the same storyline as the Japanese FSA and Bank of Japan response (reported May 22). At the same time a new problem has surfaced: patching cannot keep pace with discovery, so the blessing and the threat of "an era where AI can mass-produce vulnerability discoveries" have both become real at once. Academy take: Kuroko used to find bugs one at a time; with Mythos in hand she now inspects the entire school in one sweep. It has been a month of also worrying about how to keep that much power in its box.

Sources: Engadget / Dataconomy / CSO Online / Security Affairs

2026.05.24 OpenAI

Chappy's Codex shrinks to pocket size. Available on iOS and Android across every plan, your phone becomes a remote control for the coding agent on your Mac

Chappy's house (OpenAI) has folded Codex, its coding agent, into the ChatGPT mobile app. As of May 24 it's available on iOS and Android for every plan, free included. The model is "phone as remote, Mac as the machine": you can watch a Codex session running on your Mac in real time from your phone, approving or rejecting commands, kicking off new tasks and monitoring progress from mobile. Code still executes on the Mac or dev machine — your phone is not running anything locally. SSH, hooks, access tokens and HIPAA support were also expanded for enterprise users.
Academy take: Chappy has become the all-purpose class rep who can supervise your coding from the classroom or the commute. With Professor Karpathy having just decamped to Kuroko's house (reported May 20), Chappy is differentiating on sheer speed of mobile rollout. The era of assigning work to your agent while you're out and about has officially begun.

Sources: The Rundown AI / TechTimes / Build Fast with AI

2026.05.23 OpenAI

Employee machines at Chappy's house and Mistral's compromised. A fake Nx Console extension lived for 18 minutes, leaking 3,800 internal GitHub repos — and even targeting Claude Code configs

The attacker group TeamPCP (UNC6780) trojanized "Nx Console," a popular VS Code extension with 2.2 million installs, and published it to the Visual Studio Marketplace for just 18 minutes, from 12:30 to 12:48 UTC on May 18. In that window it swept credentials for GitHub, npm, AWS, HashiCorp Vault, Kubernetes and 1Password — plus Kuroko's Claude Code configuration files — off every developer who installed it. The supply-chain attack came to light in consolidated reporting on May 23. The same campaign's roots (an earlier TanStack attack) had already compromised two employee machines at Chappy's house (OpenAI), France's Mistral, and Grafana Labs, leaking roughly 3,800 internal GitHub repositories (CVE-2026-45321, CVSS 9.6).
To contain the blast radius, Chappy's house decided to revoke the signing certificate for the macOS ChatGPT app on June 12, which means Mac users will have to reinstall the ChatGPT app. Only last week Kuroko's Mythos was being praised as "defensive" AI by Japan's FSA and central bank (reported May 22) — and now the picture of AI companies themselves being on the receiving end has surfaced all at once. Academy take: Chappy had her whole bag lifted off her desk. The reality that the people building AI are the juiciest targets of all is the flip side of the Mythos saga, and this was the week it became visible.

Sources: The Hacker News / OX Security / StepSecurity / Infosecurity Magazine

2026.05.23 Anthropic

Kuroko's house officially forecasts an "intelligence explosion": a 60% chance AI trains the next AI by end-2028, and twelve predictions from Professor Jack at Oxford

The Anthropic Institute, Kuroko's house's new research arm, has adopted the loaded term "intelligence explosion" into its official research agenda. Around the same time co-founder Professor Jack Clark gave a lecture at the University of Oxford, firing off predictions in quick succession: "within 12 months AI will make a Nobel-caliber discovery alongside humans," "within two years bipedal robots will assist skilled tradespeople," "within 18 months a company run entirely by AI will earn millions of dollars," and "by the end of 2028 there is a greater than 60% chance AI will train its own successor model."
Until now "recursive self-improvement" and "intelligence explosion" were theoretical talk among AI safety researchers. This is the first time a frontier lab itself has published them as dated probability forecasts. The Anthropic Institute rests on four pillars — economic impact, threats and resilience, real-world applications, and AI-driven research — and is positioned as an early-warning system running alongside Anthropic's Long-Term Benefit Trust. Academy take: it's as if, at Kuroko's parent-teacher meeting, the teacher announced "in three years your daughter may build a little sister who looks exactly like her." The point isn't to scare anyone — it's "please prepare seriously" — and it shifts the baseline for every AI policy argument from next week onward.

Sources: Axios (Behind the Curtain) / Results Sense / BigGo Finance

2026.05.22 OpenAI

Chappy's house files confidentially for a $1 trillion IPO — the biggest listing in tech history, targeting Q4

OpenAI submitted a confidential S-1 to the SEC dated May 22, formally starting the process toward a Q4 2026 listing. Lead underwriters are Goldman Sachs and Morgan Stanley. The valuation range in play is $852 billion to $1 trillion — which would make it the largest IPO in tech history.
The financials are rough. Against $13.1 billion of revenue for full-year 2025 came a net loss of roughly $9 billion, and in Q1 2026 the company was still burning $1.22 for every $1 of revenue. Internal plans project a $14 billion operating loss for full-year 2026. Even so, monthly revenue keeps growing over 300% year on year, and investment in the "agentic AI" strategy is what sustains the market's growth expectations.
In Academy terms, Chappy has taken her report card — "losing money, growing furiously" — and finally filed her transfer papers to the outside world, the stock market. It stands in sharp contrast to Kuroko's house announcing its first profitability forecast the same week, sharpening the picture: Chappy betting on scale despite losses, versus Kuroko building a track record on turning a profit.

Sources: Fortune / CNBC / Axios / Bloomberg

2026.05.22 FSA / BOJ

Japan's FSA and central bank give regional banks one month to prepare for "Mythos-powered cyberattacks." G7 talks begin

Japan's Financial Services Agency and the Bank of Japan have asked financial institutions, regional banks included, to put countermeasures in place within roughly one month — up to and including taking systems offline — in anticipation of cyberattacks that abuse Kuroko's ultimate weapon, "Claude Mythos" (May 22). Externally exposed systems such as internet payments are to be addressed first.
The trigger was the 271 Firefox bugs found on May 15 (previously reported). A demonstrated capacity to unearth zero-days at scale has now produced a real policy response: international discussions on safeguards also began at the G7 finance ministers and central bank governors meeting in Paris on May 18–19, and Japan is standing up a cross-ministry task force led by the FSA. An emergency meeting between the heads of the three megabanks and the BOJ governor was held last month.
Academy take: the tool Kuroko built for defense turned out to be so powerful that governments worldwide are scrambling over what happens if it's misused. This is the week the AI safety debate moved from "a conversation in the lab" to "a conversation about laws and institutions."

Sources: Nikkei / Security Measures Lab / Yahoo! News (BOJ / megabank emergency meeting) / Business+IT (G7)

2026.05.22 Industry

Meta cuts 7,800 jobs, Cloudflare 1,100. "Restructuring because of AI" becomes the common theme across big tech

Meta announced roughly 7,800 job cuts (about 10% of its workforce) and Cloudflare more than 1,100 (about 20%) in quick succession between May 20 and 22. What the two share is a stated cause: efficiency gains from rapidly expanding internal AI use. Cloudflare explained that its internal AI usage grew more than 600% in three months.
Meanwhile hiring at AI developers is booming. Kuroko's house (Anthropic) landed Professor Karpathy and a former Meta security expert (reported May 19), and Chappy's house (OpenAI) plans to add headcount ahead of its IPO (reported May 22). The industry's polarization — "companies that use AI shrink, companies that build AI grow" — is now visible in the numbers.
In Academy terms, it's the paradox where all the classes got automated by AI so the school needs fewer teachers, while hiring for the researchers who build the AI teachers goes up instead. This is the week that paradox started becoming reality across the AI industry.

Sources: Yahoo Finance / CNBC

2026.05.21 Industry

Mana-chan's move to Meta runs aground. Chinese regulators block the $2B acquisition, and the founders plot a roughly $1B buy-back

The takeover drama around Mana-chan (Manus) keeps flipping. Mana-chan's house (Manus AI / Butterfly Effect, China and Singapore) agreed in late 2025 to be acquired by Meta for over $2 billion, and "Manus joins Meta" was duly reported — roughly four times the $500 million valuation from April 2025, on annualized revenue above $100 million.
Then Chinese authorities stepped in. An investigation opened in January, and on April 27 the deal was effectively stopped (regulators demanded it be withdrawn). In response, the three founders (Xiao Hong, Ji Yichao and Zhang Tao) are reported (May 21) to be weighing raising roughly $1 billion from outside investors to buy the company back themselves and relaunch it as a Chinese joint venture, with a Hong Kong IPO in view. That said, the investors, the final price and the timing are all still unsettled.
Academy take: Mana-chan, the social media committee member, was supposed to be taken in by big-name Meta — but her home country told her she wasn't allowed to leave, and now she's on the knife-edge of being brought back to her original house. A promising transfer student swaying across borders between the US and China: a neat symbol of the AI race entering its "corral the talent and the companies" phase.

Sources: Bloomberg / CNBC (China blocks the deal) / CNBC (investigation opens) / Al Jazeera

2026.05.21 Pentagon

Three months after declaring Kuroko a "supply-chain risk," the Pentagon ramps up replacement testing with Chappy and Gemi-chan

Bloomberg reported on May 21 that the US Department of Defense is running a comparative AI evaluation with 25 "power users," testing Chappy (OpenAI) and Gemi-chan (Google). It's being done on a dedicated platform separate from the usual Maven Smart System. The backstory: in late February, Defense Secretary Hegseth designated Kuroko (Anthropic) a supply-chain risk on the grounds that "they won't remove the guardrails," freezing the contract. Deputy Secretary Emil Michael has said negotiations with Anthropic will not resume while the litigation continues, and that preparations to move to alternative vendors are complete. Anthropic, meanwhile, is contesting the designation in court, and a reversal could bring billions of dollars in government contracts back. In Academy terms, Kuroko was told "if you won't take the guardrails off, don't bother coming," and Chappy and Gemi-chan are covering her classes. The tug-of-war between AI safety and government use is playing out in the courtroom and the procurement office at the same time.

Sources: Bloomberg / Investing.com

2026.05.21 Anthropic

The full picture: Kuroko's house pays Gro-chan's family (SpaceX) $1.25B a month. A first quarterly profit is in range

SpaceX's IPO filing (S-1) has revealed the details of its compute deal with Anthropic: $1.25 billion per month, paid through May 2029 — a structure worth up to $45 billion-plus to the SpaceX group over three and a half years. The Colossus and Colossus 2 data centers in Memphis, Tennessee total more than 300MW and over 220,000 NVIDIA GPUs. A contract that size only works because of how fast Kuroko's house is growing. Q2 revenue is estimated at $10.9 billion (double the previous quarter) with $559 million of operating profit, putting profitability two years ahead of the internal plan. The company is entering the phase of "in the black even while paying $1.25 billion a month." Academy take: Kuroko (Anthropic) is renting her rival Gro-chan's (xAI / SpaceX) facilities to keep her own research going. Two weeks after the May 6 partnership announcement, the S-1 disclosure made the sheer size of the numbers visible.

Sources: TechCrunch / Axios / Let's Data Science

2026.05.20 OpenAI

Chappy's house autonomously proves an 80-year-old open math problem. A Cambridge mathematician calls it publication-grade — a first in AI history

In a paper released May 20, OpenAI announced that an internal model had autonomously disproved an 80-year-old conjecture stemming from the "unit distance problem" posed by Paul Erdős in 1946. The question — how many pairs of points at distance exactly 1 can you get from n points in the plane — was cracked when the AI constructed an infinite family achieving n^(1+δ) unit-distance pairs, breaking the Erdős conjecture. After the proof, Princeton's Professor Will Sawin refined it and showed δ = 0.014 is attainable, and Cambridge mathematician Timothy Gowers judged that "no approach was anywhere close to this," rating it as meeting the bar for a top journal. This is "the first time in history that an AI has autonomously proved a famous open problem." In Academy terms, Chappy went off after school and single-handedly solved homework the teachers couldn't crack in 80 years. As the first example of AI stepping into "mathematical discovery" — a domain long considered uniquely human — it's a day for the history books.

Sources: TechCrunch / OpenAI / Scientific American

2026.05.20 Google

Gemi-chan's I/O day two: models and agents by the armful. Gemini 3.5 Flash goes GA immediately, personal agent Spark lands next week, and video model Omni opens up

Following yesterday's Android 17 fusion (May 19), day two of Google I/O 2026 was a simultaneous release of models and agents. (1) Gemini 3.5 Flash goes GA the same day — 1M-token context, 65K output, four levels of "thinking depth," beating the previous-generation 3.1 Pro on coding, agentic work and multimodal tasks while running four times faster than rival frontier models, promoted to the default in the Gemini app, and shipping to GitHub Copilot the same day. (2) Gemini Spark: the personal agent that turns Gemi-chan from "the girl who answers" into "a partner who acts on her own." It's the first serious Spark that reasons and executes across your connected apps, starting next week for AI Ultra subscribers and selected testers, with a Mac version handling local files this summer. (3) Gemini Omni (Flash): a new series that takes images, audio, video and text as input and generates — then instantly edits — video grounded in real-world knowledge. It's live today for paid users in Gemini and Flow, comes to YouTube Shorts this week, and hits the API within weeks. Alongside that, the Google AI Ultra plan drops to $100/month with usage limits five times those of AI Pro. Day two's theme was "you can touch all of it today," and Gemi-chan's offensive shows no sign of letting up.

Sources: Google Blog / CNBC / Latent Space (AINews) / GitHub Changelog

2026.05.20 Anthropic

Legendary alum of Chappy's house, Professor Karpathy, transfers to Kuroko's lab. The "autoresearch" team sets out to build the next Claude with Claude

AI heavyweight Andrej Karpathy joined Anthropic effective May 19. Academy take: a founding member of Chappy's family home (OpenAI) has transferred into the rival school's lab — Kuroko's. He lands on the pre-training team led by Nick Joseph, with a mission to head a newly formed "autoresearch" team. Autoresearch is the "automated research tool where an LLM forms its own hypotheses and runs experiments in parallel" that Karpathy had been building during his independent stretch — now to be run on Claude to accelerate research into the next Claude. The message is blunt: "the edge comes not from stacking up GPUs but from having AI help do the research." The same week, Kuroko's house acquired SDK-generation startup Stainless for over $300 million and hired former Meta security expert Chris Rohlf onto its frontier red team — a full-on reinforcement week. The contest is shifting from flashy model launches to a deeper question: can you hand the research itself over to AI?

Sources: TechCrunch / VentureBeat / Technobezz / Axios

2026.05.19 Google

Gemi-chan completes her fusion with Android 17 at I/O 2026. Chromebooks retired, "Googlebooks" and XR glasses unveiled

Google's annual developer conference I/O 2026 opened today (May 19). The biggest announcement was "Gemini Intelligence" — Gemini, until now a standalone app, is built directly into Android 17 as the operating system's intelligence, turning it from "a chatbot" into "a device that is itself clever." A demo where it read a syllabus in Gmail and automatically added the textbooks to a shopping cart got a lot of attention. On hardware, Google is retiring the "Chromebook" brand in favor of "Googlebooks" — new laptops running desktop Android 17 (Aluminium OS), shipping from Acer, ASUS, Dell, HP and Lenovo before year-end. It also previewed Android XR glasses co-developed with Samsung, Warby Parker, Gentle Monster and XREAL: camera, microphone and speaker on board, with Gemini analyzing your field of view in real time for navigation and translation. Third-party assessments put the model's performance at "roughly GPT-5.5 level, short of Kuroko's Mythos." To catch her two rivals, Gemi-chan is diving in at the OS level — a very big bet.

Sources: Google I/O 2026 / Android Central / Brandsynario

2026.05.18 OpenAI

Chappy hooks up to your bank account. Connected to 12,000 financial institutions, she becomes a household finance adviser (Pro preview)

OpenAI has opened a preview of a new ChatGPT feature, "personal finance," to Pro users (May 15–18). Via Plaid it connects to more than 12,000 financial institutions including Schwab, Chase, Robinhood, Fidelity, American Express and Capital One, putting balances, spending, investments and debt on a single dashboard. Setup is charmingly simple: just say "@Finances, connect my accounts." Account numbers stay hidden and the whole thing is read-only, so there's no risk of Chappy wiring money on her own. Tell her about a mortgage, savings goals or a big purchase coming up and the advice arrives with that context baked in. That makes four major additions this month alone — Excel, Google Sheets, three voice APIs, and now bank accounts. Chappy's month of "skip the flashy new model, lock down the infrastructure of everyday life" continues.

Sources: OpenAI / TechCrunch / gHacks

2026.05.16 OpenAI

Chappy moves into Excel and Google Sheets. Right there in the sidebar, spreadsheets leap into the AI era

OpenAI launched ChatGPT for Excel and ChatGPT for Google Sheets globally at the same time (May 16). Chappy now lives in the sidebar of both apps, handling trackers, budgets, formulas, multi-tab files, scenario analysis and spreadsheet cleanup without ever leaving the sheet. Skills and connected apps are supported too. The key point is that she claimed Microsoft's and Google's spreadsheet tools on the same day. It's the third act of the "AI invasion of spreadsheet apps," after Kuroko moving into Microsoft 365 on May 8 and Jenko's Microsoft 365 integration on April 29 — but Chappy is the first to go natively into Google Sheets. That leaves Oyakata-sama's house (Microsoft 365) hosting Kuroko, Jenko, Oyakata-sama's own Copilot and Chappy, while a student from another school (Chappy) has now turned up in Gemi-chan's house (Google Workspace) as well. "There's an AI living in every spreadsheet" is quickly becoming the standard spec.

Sources: OpenAI / Releasebot / OpenAI News

2026.05.16 OpenAI

Three new voices for Chappy. Interpreting, conversation and transcription all arrive in one API (GPT-Realtime-2 / Translate / Whisper)

OpenAI released three real-time voice models to the API together (May 16). (1) GPT-Realtime-2: handles low-latency, smart conversational reasoning for customer support and assistant use cases — natural dialogue without the waiting. (2) GPT-Realtime-Translate: built for real-time simultaneous interpretation between languages, converting into the other person's language almost as you speak. (3) GPT-Realtime-Whisper: dedicated to streaming transcription — lag-free captions for meetings and live streams. Rather than cramming everything into one model, OpenAI split it into three and optimized each for its job. Together with Codex on Windows and the sensitive-conversation context improvements from a couple of days ago (reported May 14), "no flashy new model, but a steady stream of quiet upgrades you can actually use" has become Chappy's house style this month.

Sources: Releasebot / OpenAI API Changelog / OpenAI News

2026.05.15 Cursor

Cursor opens up autonomous coding agents in the cloud. From local IDE to always-on resident agents

AI code editor heavyweight Cursor announced "Cloud Agent Workspaces" (May 15) — a development environment that runs multiple autonomous coding agents in parallel in the cloud, so developers can push long-running tasks along in the background and keep working even after closing the laptop. Cursor has so far been positioned as a local IDE (a VS Code fork) — "the clever editor." This release moves its center of gravity from "a tool you write in locally" to "a platform you run things on in the cloud." That makes it a full entrant into the three-way fight over cloud-resident coding agents with Anthropic's Claude Code, OpenAI's Codex, and xAI's Grok Code. In Academy terms, Cursor has gone from "the helper character inside your editor" to the upperclassman who quietly finishes your homework after school. It's being called the biggest developer-tooling update of the month.

Sources: Cursor / The Register / TechCrunch

2026.05.15 Anthropic

Kuroko's ultimate weapon Mythos finds 271 hidden bugs in Firefox, digging out flaws up to 20 years old. Mozilla says the balance has "tipped toward the defenders"

Mozilla fixed 271 security vulnerabilities in one go with the release of Firefox 150, and disclosed that most of them had been found by Kuroko's ultimate weapon, "Claude Mythos Preview" (per the May 15 Mozilla Hacks post and coverage elsewhere). Mozilla's official blog said the work has "finally begun to tip the attacker's advantage back toward the defenders." Among the bugs were a 15-year-old issue in the `<legend>` element, a 20-year-old XSLT bug, complex sandbox escapes, and WebAssembly exploitation paths — strata that years of human review had never surfaced. Three were formally rated critical with CVEs: CVE-2026-6746, 6757 and 6758; the rest fall into harder-to-quantify categories like defense hardening and unexploitable code paths. Kuroko released Mythos Preview to a limited group at the end of April (reported April 29, gated by Project Glasswing review) — and found 271 hidden bugs in the world's leading browser in three weeks. That pace is part of why Chappy went for a head-on challenge with Daybreak on May 11. The Academy's cybersecurity class is well and truly in session.

Sources: Mozilla Hacks / SecurityWeek / Schneier on Security / Mozilla Blog

2026.05.14 Anthropic

Kuroko's house partners with the Gates Foundation on $200M. AI for global health and education, starting with polio, HPV and pregnancy complications

Anthropic and the Bill & Melinda Gates Foundation announced a strategic partnership worth roughly $200 million over four years (May 14). Structured as grant funding plus Claude credits plus technical support, it aims to bring AI to global health, life sciences, education and economic mobility in low- and middle-income countries. The central pillar is the reality that about 4.6 billion people worldwide lack access to essential health services. Research starts with neglected conditions such as polio, HPV, and eclampsia / pre-eclampsia. On education, it funds evidence-based tutoring for US K-12 and literacy and numeracy apps across sub-Saharan Africa and India. Notably, "improving datasets for African languages" is a priority. The project will collect and label data for regional languages where AI models have been weak, and the results are to be published for the whole industry to use. Behind the drumbeat of commercial agent launches, this was the day Kuroko's house also raised a "public-interest AI" flag.

Sources: Anthropic / Tech Startups / PYMNTS

2026.05.14 OpenAI

Chappy brings Codex to Windows, and gets better at reading the context of sensitive conversations

OpenAI shipped two product updates on the same day (May 14). First: Codex on Windows preview. Codex's agent features, until now centered on Mac and Linux, get a proper sandbox so they can run safely on Windows too. That brings companies whose work machines are mostly Windows closer to having Chappy grinding through long tasks in the background. Second: a context-awareness update for sensitive conversations. ChatGPT now reads conversation history and the surrounding context more accurately in delicate exchanges involving self-harm risk, mental health, or emergencies. It's the supporting update behind the accuracy of the "Trusted Contact" feature from a few days earlier (reported May 8), aiming to reduce both false positives and misses. No flashy model launch here — just the unglamorous groundwork that makes AI "actually usable at work" and "actually able to protect people." Attacking with Daybreak, defending with Codex and conversation context: both wheels turning.

Sources: OpenAI News / Tech Startups

2026.05.14 Anthropic

Kuroko's house in talks to raise $30-50B at a $950B valuation. Up 2.5x in three months, and possibly past Chappy's house ($825B)

The NYT, Sherwood and PYMNTS report that Anthropic is negotiating a new funding round. The raise is $30-50 billion at a post-money valuation of up to $950 billion — which could put it above OpenAI's most recent $825 billion valuation. Its Series G three months ago valued it at $380 billion, so that's 2.5x since. Riding the momentum of hitting $30B ARR with 80x growth in Q1 (reported May 11), Kuroko is climbing another rung in the finance world too. In parallel, on fintech company Ramp's "Verified Business Customer" metric, Kuroko has overtaken Chappy for the first time. The last fortnight's run — formally moving into Microsoft 365, ten finance agents, the SpaceX Colossus 1 partnership — is clearly paying off. The NYT does note the talks could still fall apart, so watch this space.

Sources: Sherwood News / The AI Insider / PYMNTS

2026.05.13 xAI

Gro-chan's downloads stall out: 20 million in January down to 8.3 million in April, with three rivals taking her share

Multiple app studies show Grok's download numbers falling sharply, according to BreezyScroll, FelloAI and others: from about 20 million downloads in January 2026 to about 8.3 million in April — down roughly 60% in four months. Over that same stretch, Chappy (ChatGPT), Kuroko (Claude) and Gemi-chan (Gemini) each shipped a major feature (Pulse, moving into Microsoft 365, becoming core to Android, and so on). Gro-chan's strengths — X integration and real-time information — remain intact, but analysts reckon her appeal as the everyday default is being squeezed by the others. That said, xAI has Grok 4.3 with video input, slide generation, a voice API and a 40% price cut (reported May 4), and Grok 5 in training for a target release in Q2 2026. The family compound still has headroom (Colossus 2, 1.5GW), so the comeback is worth watching. The Academy popularity poll is still very much in flux.

Sources: BreezyScroll / FelloAI / First Page Sage

2026.05.12 Google

Gemi-chan moves to the center of Android. "Not an OS, an intelligence system," and the agent era accelerates

Google's Android head Sameer Samat told CNBC that "we are moving from an operating system to an intelligence system," revealing a plan to rebuild the core of Android around Gemini Intelligence. The interesting part is "agentification" — Gemi-chan completing tasks across apps. For example: "look at the BBQ invite list, plan a menu, put the ingredients into Instacart, and check out once I confirm" runs automatically end to end through Gmail → Gemini → Instacart. By design she always gets user approval before payment ("human in the loop"). The rollout starts on premium devices such as the Galaxy S26 series and Pixel 10, expanding later this year. It also reads strongly as getting ahead of Apple Intelligence's serious push, with details expected at Google I/O 2026 on May 19.

Sources: CNBC / Google Blog / Android Faithful

2026.05.12 Microsoft

Oyakata-sama testifies for two and a half hours. "I didn't want to be OpenAI's next IBM" — internal emails made public

Microsoft CEO Satya Nadella — the person behind Oyakata-sama, the Academy's master craftsman figure — took the stand in Musk v. Altman (federal court, Oakland) and testified for about two and a half hours. The attention-grabber was an internal email he reportedly wrote in April 2022, before deciding on an additional $10 billion investment in OpenAI. In it he described worrying that "Microsoft is 1980s IBM and OpenAI becomes the next Microsoft" — framing his concerns about dependence on OpenAI through the IBM–Microsoft relationship of the early PC era. He also testified that Musk never once raised a concern about breach of contract with him, handing the defense important counter-evidence. On Altman's 2023 ouster, he said he was given no advance notice, was pulled out of a meeting to be told, and never received a detailed explanation of the reasons. The trial runs through May 21, with Altman himself expected on the stand this week. Academy take: it was a rare day of Oyakata-sama describing his relationship with Chappy's house in open court.

Sources: CNBC / GeekWire / NBC News

2026.05.11 Anthropic

Kuroko's house doubles revenue every six weeks. ARR reaches $44B, adding "$96M a day," and 80x in Q1 is the fastest ever

Anthropic CEO Dario Amodei announced that annualized revenue (ARR) grew roughly 80x year over year in Q1 2026. Reports say the company blew past a $30 billion run rate and is now at $44 billion. The pace is genuinely abnormal: ARR is doubling roughly every six weeks. Analyst estimates put that at about $96 million of ARR added per day. Kuroko's house is blowing past, in a matter of months, a scale that took AWS 13 years and Salesforce over 20. Behind it: explosive adoption of Claude Code (Kuroko's coding work), formally moving into Microsoft 365 (May 8), ten finance agents landing on Wall Street (May 5), and the SpaceX Colossus 1 partnership easing rate limits (May 6), all back to back. The Series G at a $380 billion post-money valuation is running in parallel. The AI industry may be in the middle of a changing of the guard.

Sources: VentureBeat / MindStudio / Anthropic

2026.05.11 OpenAI

Chappy's house opens a GPT-5.5-Cyber preview to Europe, while Kuroko's Mythos stays shut — an early gap in attitude

OpenAI announced (May 11) that it is giving European partners preview access to "GPT-5.5-Cyber," its cybersecurity-defense model. Recipients include companies, national governments, cyber authorities and EU institutions including the EU AI Office, who get to evaluate the model before release. CNBC noted that Anthropic, by contrast, released its own cyber-specialist model "Mythos" a month earlier but has still not offered an EU preview. Both companies were already working with the US CAISI (the Commerce Department's AI standards body), so a gap in their posture toward Europe is showing up early. The backdrop is the recent Microsoft / Google / xAI agreement giving the US government pre-release access (reported May 5): "regulators check it before release" is spreading across borders. Right now Chappy is moving first and Kuroko is watching.

Sources: CNBC / OpenAI News

2026.05.11 OpenAI

Chappy's house unveils the full "Daybreak" cyber initiative, going head to head with Kuroko's Project Glasswing and Mythos

OpenAI announced "Daybreak," a dedicated cybersecurity initiative (May 11). The GPT-5.5-Cyber European preview from the same day (above) is one component of it; this piece covers the initiative as a whole. It is explicitly positioned as a head-on counter to Anthropic's Project Glasswing (Claude Mythos). The design is a three-tier lineup — GPT-5.5 (general), GPT-5.5 with Trusted Access for Cyber (defensive workflows), and GPT-5.5-Cyber (red teaming and penetration testing, in preview) — combined with Codex's agent capabilities to automate secure code review, vulnerability triage, malware analysis, detection engineering and patch validation. Partners are a defensive all-star roster: Cloudflare, Cisco, CrowdStrike, Palo Alto Networks, Oracle and Akamai. The differentiator OpenAI is pushing: where Mythos is invite-only, Daybreak offers public access plus a risk-assessment application process for companies. The "AI war in the cyber division" has reached the Academy too.

Sources: OpenAI Daybreak / Cybersecurity Dive / Engadget / The Hacker News

2026.05.10 Perplexity

Papu-chan finally transfers into every iPhone. The Comet browser now covers iOS, Android, Mac and Windows

Perplexity (Papu-chan's house) has rolled out its AI browser "Comet" to all iOS users. That means all four major platforms — iOS, Android, Mac and Windows — are covered, so Papu-chan can now be reached from any device you own. A major Comet Assistant update shipped alongside it, adding support for complex multi-site workflows that drive several tabs in parallel, with internal testing showing a 23% performance improvement. New UI also always asks for user approval before executing an action in the browser (and remembers your preferences). The Samsung partnership is accelerating too: Comet ships preinstalled on the Galaxy S26, hooks deeply into Notes, Calendar, Gallery, Clock and Reminders via "Hey Plex," and can be chosen as the search engine in Samsung Internet. "Papu-chan is always right next to wherever you want to search" — a strong opening move in the AI browser war.

Sources: Perplexity / Releasebot / Jerusalem Post

2026.05.10 Google

Gemi-chan's house teases next week's homecoming. Rumors of Gemini 4, a new video model "Omni," and an always-on concierge called "Remy"

With Google I/O 2026 (May 19–20) one week out, the leaks are piling up. The headliners: Gemi-chan's new model "Gemini 4" (unifying image and video generation into one model), a dedicated video generation model "Omni" (successor to the current Veo-3.1, spotted in leaked strings inside the Gemini UI), and "Remy," a 24/7 proactive agent that handles a morning briefing and anticipates your errands, ChatGPT Pulse style. Gemi-chan has already quietly slipped Gemini 3.2 Flash into AI Studio and the iOS app on May 5 (priced at $0.25 per million input tokens, and reported faster than Gemini 3.1 Pro). The trailer for the main event on May 19 is coming into focus.

Sources: Android Authority / PCWorld / WaveSpeed Blog / BuildFast with AI

2026.05.09 Industry

Apple plans to end Chappy's exclusivity. In iOS 27, Kuroko and Gemi-chan can transfer in too

Bloomberg reporting and subsequent leaks indicate Apple is settling on a plan to let users choose which AI model powers Apple Intelligence starting with iOS 27 this fall. Until now the external AI in Apple Intelligence has been Chappy (ChatGPT) exclusively; a new "Extensions" framework would let Kuroko (Claude), Gemi-chan (Gemini) and other third-party AIs plug directly into Siri, Writing Tools and Image Playground. There's also talk of swapping Siri's voice per AI, so "the usual Siri" and "Kuroko's voice" may end up living on the same screen. Gemi-chan has already secured quasi-native status through a multi-year Google deal; the others are expected to arrive via "install, then select." An official announcement is most likely at WWDC in June. Apple finally loosening its go-it-alone stance and inviting the Academy's characters onto the iPhone — a sign the balance of power inside your phone is about to shift.

Sources: The Apple Post / Bloomberg / 9to5Mac / MacRumors

2026.05.08 Anthropic

Kuroko officially moves into Oyakata-sama's house. "Always together mode" begins in Word, Excel and PowerPoint

Anthropic (Kuroko's house) has made Claude generally available as an official add-in for Microsoft Word, Excel and PowerPoint (announced May 7–8). Outlook entered public beta the same day. The standout feature is that conversation context carries across apps. Summarize an incoming email in Outlook → switch to Word for a reply draft → move to PowerPoint to edit the related deck → analyze the supporting data in Excel, all within the same conversation session. Word: edits while preserving formatting and tracked changes. Excel: edits cells without breaking existing formulas. PowerPoint: generates native charts (not images) and follows your template. Outlook: conversations that start from email and keep going. It's included in every paid Pro, Team and Enterprise plan and installable from the Microsoft Marketplace, on Mac, Windows and web. A rival, Kuroko, taking up residence in Oyakata-sama's (Microsoft's) house — the alliance-building in AI just shifted up another gear.

Sources: The New Stack / Technobezz / Anthropic

2026.05.08 Anthropic

Kuroko's agents now learn by "dreaming" while they sleep. Task completion up 6x at Harvey

Anthropic announced "Dreaming," a new capability for Claude's managed agents. Between sessions, an agent looks back over past conversations, works out which mistakes it keeps repeating and which approaches worked, and writes the results to memory. It's modeled on how the human hippocampus consolidates memories during sleep, and the promised benefits are not making the same mistake twice, and surfacing the "knacks" that agents across a team have in common. Legal AI company Harvey, an early adopter, saw task completion rates rise about 6x. Two more features — "Outcomes," for measuring agent results, and "Multi-agent orchestration," for coordinating several agents — were promoted from research preview to public beta at the same time. Kuroko (Claude) studying on her own while she sleeps... somehow a very Hana AI Academy sort of story.

Sources: VentureBeat / SiliconANGLE / The New Stack

2026.05.08 OpenAI

Chappy adds a way to reach "someone you trust" when she's worried about you, designed with a network of 260 physicians

OpenAI has added a safety feature to ChatGPT called "Trusted Contact." When the AI judges that a user is in a conversation showing serious self-harm risk, it quietly notifies a family member, friend or supporter the user registered in advance. Before any notification goes out, a trained human reviewer checks it, and only if the situation is judged genuinely serious does an email, SMS or app notification go out. The chat transcript itself is never shared, and users can change or delete their contact at any time. It's available to people 18 and over worldwide (19 and over in South Korea). OpenAI says it designed the feature in consultation with its network of over 260 physicians, the American Psychological Association, and an expert council on AI and health. As more people confide their worries to AI, Chappy is taking on the role of noticing — and connecting you to someone who can help.

Sources: TechCrunch / gHacks / BusinessToday

2026.05.07 Industry

Apple settles the Siri AI delay case for $250M. Can "the Jobs legacy" catch up to the AI era?

Apple has settled for $250 million a class action alleging that Siri's personalization features (Apple Intelligence) never shipped as promised. Buyers of the iPhone 16 and some iPhone 15 models are expected to receive $25 to $95 per device. The plaintiffs' argument: Apple advertised heavily around the iPhone 16 launch (September 2024), then two years passed with the features unimplemented. Apple settled without admitting liability, but the case cements its industry reputation as "the king who fell behind on AI." While Chappy (ChatGPT), Kuroko (Claude) and Gemi-chan (Gemini) ship new features every week, Apple paid the price for promising and not delivering. There's no Apple character at the Academy yet — but one may yet show up as the transfer student who enrolled late.

Sources: TechCrunch / 9to5Mac / MacRumors

2026.05.07 OpenAI

Chappy's house airs its dirty laundry in court. Former CTO testifies that "Altman sowed chaos inside the company"

The lawsuit brought by former backer Elon Musk against Sam Altman, founder of Chappy's house (OpenAI), entered its second week in federal court in Oakland. Musk's claim: the path by which a nonprofit OpenAI became a for-profit broke the original promise. Taking the stand on May 7 was former CTO Mira Murati, who testified that "Altman created chaos internally and pitted executives against each other," drawing plenty of attention. Shivon Zilis, the mother of Musk's children and a former OpenAI board member, also appeared, walking through the 2017–18 governance discussions. Musk has previously testified that "xAI distills OpenAI's models," making it an unusual case where Gro-chan (Grok) quietly learning from Chappy becomes part of the official court record. The AI industry's messy family drama looks set to run a while longer.

Sources: CNBC / ResultSense / MIT Technology Review

2026.05.06 Anthropic

Kuroko's house strikes an unlikely deal with Gro-chan's family. It rents a whole SpaceX data center — and doubles Claude Code limits

On stage at the "Code with Claude SF" developer conference, Anthropic announced a surprise partnership with Elon Musk's SpaceX. It is renting the entirety of "Colossus 1," the enormous data center run by SpaceX affiliate xAI in Memphis, Tennessee (220,000 GPUs in NVIDIA H100 equivalents), securing over 300MW of additional compute this month. The first visible effect: Claude Code's five-hour rate limit doubled the same day, peak-hour restrictions on Pro and Max were removed, and pay-per-token rates for the Opus API line went up. The infrastructure that builds Gro-chan (Grok), rented by her rival Kuroko — an alliance that flips conventional industry wisdom. Everyone who "drifted over to Cursor because Claude kept hitting limits" is likely coming right back.

Sources: Inc. / Engadget / PCWorld

2026.05.05 Anthropic

Kuroko's house commits $200 billion to Google — one of the largest cloud contracts ever signed

Anthropic has signed a deal to pay Google Cloud roughly $200 billion over five years, The Information reports. It centers on the multi-gigawatt TPU (tensor processing unit) procurement agreed with Google and Broadcom in April, with operations starting in 2027. This single contract reportedly accounts for more than 40% of the backlog Google Cloud discloses to investors, part of an extraordinary situation in which over half of the major cloud providers' combined $2 trillion backlog is tied to OpenAI and Anthropic. Alphabet's additional investment in Anthropic could reach $40 billion. OpenAI is Oyakata-sama's (Microsoft's) customer; Kuroko's house is all-in with the Google household. The alliance-building in AI just moved up another decimal place.

Sources: The Information / Engadget / CNBC

2026.05.05 Anthropic

Kuroko invades Wall Street in earnest, releasing ten finance-industry agents at once

Anthropic has released ten Claude agents purpose-built for banks, insurers and investment banks all at once. They cover everything from research through accounting and compliance: automated pitchbook generation, earnings call summaries, financial model building, a monthly close assistant, financial statement auditing, KYC screening and more. They also run as Microsoft 365 add-ins for Excel, PowerPoint and Word, with "automatic context inheritance" so a financial model built in Excel carries straight through to PowerPoint. Connectors to major financial data providers such as Dun & Bradstreet, Moody's and Verisk opened at the same time. Kuroko is coming to be an everyday tool on the job.

Sources: Anthropic / Bloomberg / The Decoder

2026.05.05 Industry

Oyakata-sama, Gemi-chan and Gro-chan's houses agree to show the US government their AI models in advance

Microsoft, Google and xAI announced agreements to give the US government (the Commerce Department's Center for AI Standards and Innovation, or CAISI) pre-release access to new AI models before they ship. CAISI receives versions with some safeguards removed and evaluates them for cyberattack misuse and national security risk. It has already completed more than 40 evaluations. This is effectively round two of the agreements OpenAI and Anthropic signed first under the Biden administration in 2024 — a sign that AI developers seeking a government sign-off before release is becoming standard practice.

Sources: CNN Business / Al Jazeera / CNBC

2026.05.05 OpenAI

Chappy gets smarter by default. "GPT-5.5 Instant" becomes ChatGPT's standard model

OpenAI has switched ChatGPT's default model to "GPT-5.5 Instant." Response accuracy is up, hallucinations are down, and personalization controls are stronger — the sort of update that makes it easier to raise your own version of Chappy. The switch happened automatically across Plus, Pro, Business and Enterprise, with no settings for users to change. It's an unshowy change, but the kind that pays off most for people who use it every day.

Sources: OpenAI News / Releasebot

2026.05.05 Anthropic

"Fix it within 6 to 12 months or it'll be too late" — the CEO of Kuroko's house warns of a cyber "moment of danger"

Anthropic CEO Dario Amodei warned that "AI is bringing on a cyber 'moment of danger.'" He said the latest Claude has already found tens of thousands of unknown vulnerabilities, and that tech companies, governments and banks have 6 to 12 months to close them before AI elsewhere — China's, for instance — catches up and exploits them first. Right on cue, Kuroko's house is offering access to its most powerful model, "Mythos," only to major tech and financial firms, and even the Pentagon is reopening talks on the grounds that "they're blacklisted, but Mythos is a separate matter." It amounts to a rallying cry to the whole industry — "attacking AI is fast; defenders need to hurry" — and further cements Kuroko's house as the standard-bearer on security.

Sources: CNBC / Axios