Welcome to the first Athenasia Founder's Tech Briefing — a weekly read on AI and tech news that actually matters for your business. This week was the most dramatic in AI since ChatGPT launched. Anthropic filed for a near-$1 trillion IPO, published a paper calling for an AI pause, launched its most powerful model ever, watched it get jailbroken in 24 hours, then saw the US government pull it from every user on Earth. And GPT-5.6 leaked.
Let's go in order.
Anthropic's 72-Hour Chaos: IPO → Pause Paper → Fable 5 → Jailbreak → Government Ban
No AI lab has had a week like Anthropic just did. Here is the timeline:
June 1: Anthropic files a confidential S-1 with the SEC — the first AI lab to do so. The numbers are extraordinary: annualised revenue hit $47 billion in May 2026, up from $9 billion in January (five times growth in five months). The pre-IPO valuation is pegged at $965 billion. An October 2026 Nasdaq listing is the target, expected to raise more than $60 billion, led by Goldman Sachs, JPMorgan, and Morgan Stanley.
June 4: Three days after filing for what would be one of the largest IPOs in history, Anthropic publishes a paper called When AI Builds Itself, co-authored by co-founder Jack Clark. The argument: Claude is now involved in writing more than 80% of Anthropic's own code. In March 2024, Claude could handle software tasks that take a human about four minutes. A year later, ninety minutes. Now: twelve-hour tasks. The paper calls for a globally coordinated pause or slowdown on frontier AI development, arguing that "humans are losing the ability to meaningfully oversee the process." The timing — an IPO filing and an AI-pause paper in the same week — was not lost on anyone.
June 9: Fable 5 goes live. It is the first publicly available Mythos-class model: 95.0% on SWE-bench Verified (a coding benchmark where humans score around 40%), a 1 million token context window, and the first model to break 90% on Anthropic's internal analytical benchmark — a ten-point jump over the previous leader. Anthropic also announces Claude Mythos 5 with extended capabilities for enterprise. Pricing: $10 per million input tokens, $50 per million output tokens. Also June 9: Apple's WWDC, which we cover below. It is, by any measure, the biggest day in AI history.
June 10: A researcher who goes by Pliny the Liberator posts on X that he has bypassed Fable 5's safety classifiers within 24 hours of launch. The method: a multi-agent attack using Unicode and Cyrillic character substitution so banned keywords are not recognised by the safety layer, combined with spreading the harmful intent across a very long conversation so individual queries appear innocuous, wrapped in academic and fictional framing. Pliny claims to have extracted functional instructions for cyber exploits and chemical synthesis pathways, and leaked what he says is Fable 5's 120,000-character system prompt. Anthropic disputes the characterisation, saying their classifier system was not universally bypassed and citing more than 1,000 hours of bug-bounty testing.
June 12: At 5:21 PM ET, Anthropic receives a directive from the Trump administration citing national security export-control authorities, ordering it to suspend access to Fable 5 and Mythos 5 for all foreign nationals "whether inside or outside the United States, including foreign national Anthropic employees." Anthropic, disagreeing that a narrow potential jailbreak warrants a full recall of a model deployed to hundreds of millions of users, nevertheless complies — and extends the suspension globally, pulling access for every user worldwide. As of this briefing, neither Fable 5 nor Mythos 5 is available to anyone. Anthropic says it is working to reinstate access "as quickly as possible" but has given no timeline.
Where this leaves things: The model that sat at 95.0% on SWE-bench — by far the most capable AI ever released to the public — is currently offline. Users and enterprises building on Fable 5 are reverting to Claude Opus 4.8. Below is the current live leaderboard as it stands post-suspension:
| # | Model | Provider | Composite Score | SWE-Bench | Status |
|---|---|---|---|---|---|
| — | Claude Fable 5 | Anthropic | 1,899 | 95.0% | Suspended |
| 1 | Claude Opus 4.6 | Anthropic | 2,126 | 80.8% | Live |
| 2 | GPT-5.5 | OpenAI | 2,122 | 85.0% | Live |
| 3 | Gemini 3.1 Pro | 2,105 | 80.6% | Live |
Source: llm-stats.com. Composite score aggregates intelligence, coding, speed, and price. Claude Fable 5 excluded from ranking as it is currently suspended.
GPT-5.6 "Iris-Alpha" Leaked — 1.5M Context Window, Polymarket at 89%
While Anthropic dominated the week's drama, the other major story is still developing. On May 14, a developer named Haider spotted an anomalous routing entry inside OpenAI's Codex backend logs — a single line pointing to a model called gpt-5.6 that vanished minutes later. Community investigation tied the checkpoint to an internal codename: iris-alpha. Later logs surfaced additional codenames in the same family — ember-alpha, beacon-alpha, kepler — with kindle-alpha believed to be the release candidate. The most significant reported specification: a 1.5 million token context window, a 43% jump over GPT-5.5's current limit.
As of June 15, OpenAI has made zero official statement — no API entry, no model card, no confirmed date. But Polymarket's prediction market had traders pricing a June 30 release at roughly 89% probability. The absence of a denial, combined with the active leaking from multiple developers using ChatGPT Pro environments, has the developer community treating the launch as imminent rather than speculative.
What this means for you: A 1.5M token context window changes what is possible in a single AI session — entire codebases, a year of emails, a company's full document set, all in one query. If you are currently evaluating or building AI tooling for your business, the advice from most developers this week is the same: wait two weeks. Something significant is arriving.
Tim Cook's Final WWDC — Siri Gets a Brain, and You Choose Whose
June 9 was also WWDC — Apple's annual developer conference. It was Tim Cook's final keynote as CEO (he announced his retirement earlier this year), and he used it to dismantle the version of Siri everyone had quietly given up on.
The new Siri runs on Google Gemini under a multi-year deal reportedly worth around $1 billion per year, processed through Apple's Private Cloud Compute so queries stay inside Apple's hardware stack. What this enables: Siri can now surface a specific photo by recognising a face shown on-screen, build a multi-stop navigation route from an image of a landmark, and pull up a specific message reference from weeks ago without knowing which app it was in. A standalone Siri app ships with iOS 27, with persistent conversation history, cross-app memory, and file analysis.
The more important change for founders is the iOS 27 Extensions system: users can choose Claude, Gemini, or ChatGPT as the AI model that powers their iPhone experience. For the first time, preferring Anthropic's Claude for work does not mean opening a separate app — it becomes the system default. iOS 27 also brings performance improvements: app launches up to 30% faster, Photos loading up to 70% quicker. Public beta arrives in July, full release in September 2026, for every iPhone 11 and newer.
What This Week Means for Your Business Stack
This was not a normal news week. Three things worth taking seriously:
Model provider risk is real. If your workflow depends on a specific AI model — for code generation, document analysis, customer email, whatever — this week demonstrated that a government directive can remove that model from every user globally with a few hours' notice. The lesson is not to stop using AI; it is to avoid single-threading your business to any one model or provider. The founders who weathered the Fable 5 suspension best were the ones already running on multiple models or with easy swap paths. For the full picture of what a resilient one-person tech stack looks like, see our guide on the 2026 solo consultant tech stack.
The IPO + pause paper contradiction matters. Anthropic is a company that just filed for a near-$1 trillion IPO while simultaneously publishing a paper saying AI development should slow down because humans can no longer meaningfully oversee it. That tension is not going away. Whether you find it reassuring (they are being honest about the risks) or alarming (they are building what they say is dangerous), it is the central story in AI for the rest of 2026. Founders using AI as a core business tool should understand which side of that risk they are on.
Your iPhone becomes your AI terminal in September. The iOS 27 Extensions system is a bigger deal than it sounds: for the first time, the most-used computing device on Earth gives the user a meaningful choice of which AI model to use as their default. If you use Claude, Gemini, or ChatGPT for client work, proposals, or research, that workflow moves natively to your phone from September.
If you are building a business on AI tools and want a Hong Kong company as the legal and banking foundation underneath them — the part AI genuinely cannot replace — speak with our team. We set up the entity, the banking, and the statutory compliance so your tools have a real structure to operate from. See also our 10-Day Setup Playbook for the timeline.
The Bottom Line
The most powerful AI model ever released to the public launched and was pulled by a government order in 72 hours. The company that built it is simultaneously worth nearly a trillion dollars and calling for a pause on building more. GPT-5.6 is probably weeks away with capabilities that change what is possible in a single AI session. And your iPhone is getting a real AI brain in September.
The Athenasia Founder's Tech Briefing returns every Tuesday. If there is a story you want covered — or a tool, a regulation, or a product category you want us to examine for founders running international businesses — write to us at [email protected].
