Editor's note. This is the rare day where the news cycle is genuinely dominated by a single government-enforcement event, so it takes the lead even though it is a regulatory/state-action story — the topic-balance guidance permits exactly this, and the event in question is the most direct illustration of the sovereignty thesis the brief has run all year. A US export-control directive abruptly pulled Anthropic's two most capable models from every customer worldwide. The other two mains are deliberately non-regulatory: the open-weight cohort finally delivering (Open Source) and the AI build-out's debut on public markets (Enterprise/markets). Regulation is otherwise held to a single quick hit. Saturday's run did not fire, so this brief covers roughly 48 hours.
A US directive pulls the two most capable Claude models from every customer worldwide
On Friday evening Anthropic received a government order that should concentrate the mind of anyone outside the United States who builds on frontier AI. Citing national-security authorities, the US government issued an export-control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national — inside or outside the country, including Anthropic's own foreign-national employees. Because Anthropic cannot cleanly separate foreign from domestic users at that granularity, the net effect, in its own words, is that it "must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance." The company says it received the letter at 5:21pm ET on 12 June and that access to its other models is unaffected. The trigger, as reported by Time and TechRadar, was a jailbreak demonstration that surfaced earlier in the week; Anthropic counters that the bypass produced only "minor findings" already reproducible with publicly available models such as GPT-5.5, and warns that recalling a commercially deployed model over a narrow, non-universal jailbreak "would essentially halt all new model deployments for all frontier model providers."
Strip away the dispute over whether the safety concern was warranted, and the structural fact is the one that matters here. The most capable commercial AI in the world was switched off for every European, Asian, and other non-US customer overnight, by a unilateral act of a single government, with no notice and no appeal. This is not the CLOUD Act's quiet data-access risk in the abstract — it is the live demonstration. A European bank, hospital, or ministry that had standardised on Mythos 5 woke up on Saturday to a model it could no longer call, for reasons it was not party to and cannot contest. Whatever one thinks of the export-control rationale, dependence on a frontier model that a foreign executive branch can revoke by directive is a continuity risk that no SLA covers. The mitigation is not a louder vendor promise; it is infrastructure whose availability is not contingent on another state's national-security posture — open weights you hold, and a routing layer that can fail over to a model nobody can switch off remotely. (heise coverage.)
The open-weight counterweight finally lands — twice in one week
If the lead asks what you fall back to when a frontier model is pulled out from under you, this week supplied two answers. MiniMax's long-overdue M3 weights are now live on Hugging Face — the official 428B-parameter mixture-of-experts repository plus community quantisations (GGUF, MXFP8, NVFP4) — closing the credibility gap that had widened each day the release slipped past its ~11 June deadline. Two days earlier Moonshot AI shipped Kimi K2.7-Code under a Modified MIT licence, a one-trillion-parameter (32B active) coding model that the company reports cuts reasoning-token usage roughly 30% versus its predecessor while posting double-digit gains on its own coding benchmarks. Both run today on the open inference stack — vLLM, SGLang, llama.cpp — meaning a European deployer can pull either model into a Norwegian or German data centre this afternoon.
The caveat that has followed M3 since launch still applies and is worth restating plainly: the headline numbers — 59% on SWE-Bench Pro, ahead of GPT-5.5 and Gemini 3.1 Pro — were produced by MiniMax on its own infrastructure with baselines it chose, and independent reproduction on neutral boards is only now beginning. Kimi's figures carry the same asterisk. But the point of an open-weight model is not that you take the vendor's benchmark on faith; it is that you can download it, run your own evaluation on your own data, and keep running it regardless of what any government or any pricing committee decides next week. That property — not a leaderboard rank — is what made this week's twin releases the natural reply to Friday's directive.
The AI build-out's biggest day on public markets
SpaceX completed the largest IPO in history and then some: SPCX closed its first session up 19.3% at $161.11, having raised roughly $75bn by selling more than 555 million shares at a $135 offer price, and touched a valuation above $2.25tn intraday before settling near $2tn. The retail-heavy book — orders well past $250bn, with shares jumping on debut — capped a fortnight in which the capital structure of the AI frontier moved decisively into public hands, alongside confidential S-1 filings from both OpenAI ($852bn mark) and Anthropic (last raised at $965bn).
Read against this weekend's lead, the two stories triangulate uncomfortably. The supply of frontier AI is increasingly underwritten by US-listed entities answerable to quarterly demand, and the access to it is governed by US national-security authorities that can be exercised without warning. Both levers — the capital that builds the models and the state that can switch them off — sit in the same jurisdiction, upstream of every enterprise's AI spend. For a regulated European buyer the conclusion is not anti-American; it is simply prudential. Concentration risk that spans both your supplier's balance sheet and its government's export policy is the kind a continuity plan is supposed to diversify. Portability, the option to host open weights yourself, and a provider-agnostic routing layer are the unglamorous instruments that turn a single point of failure into a managed dependency.
Quick Hits
- The price war gains a third front — even as OpenAI weighs deep token-price cuts to blunt Anthropic ahead of dueling IPOs, Chinese open-weight models are reported to be undercutting both US labs by as much as 9x on cost — and this week's M3 and Kimi K2.7-Code releases put real downloadable weight behind that pressure rather than just an API quote.
- Microsoft keeps building its own bench — following the MAI line it unveiled at the start of June, Microsoft and Google are both pushing first-party coding models to lessen their reliance on OpenAI and Anthropic — the commoditisation of the coding tier continues from both the open and hyperscaler directions.
- Gemini 3.5 Pro still hasn't shipped — Google's promised June GA (2M context, Deep Think, a ~$15/$60 per-million read) remains in limited Vertex preview with two-and-a-half weeks left in the month; the slip keeps the top of the price ladder unsettled just as the Fable/Mythos suspension removes two premium options from the table entirely.
- EU AI Act clock, one regulation item — GPAI obligations are now 49 days out (2 August 2026), and the Digital Omnibus postponement agreed politically on 7 May (high-risk Annex III to Dec 2027, AI-in-products to Aug 2028) still awaits formal adoption and Official Journal publication before that date — provisional until it lands.