Editor's note. Today's three stories form one argument about who gets to use a frontier model. The lead is the cleanest signal yet that the closed US frontier is being walled off: OpenAI shipped GPT-5.6 but, at the government's request, only to roughly twenty vetted "trusted partners." The second story is part of why that wall is going up — Anthropic's accusation that Alibaba siphoned 28.8 million queries out of Claude to train a rival. The third is the structural escape hatch for everyone outside the wall: open-weight models you can't be locked out of. Regulation appears only in the quick hits, so there are zero regulation mains today.
OpenAI ships its strongest model yet — and the US government decides who gets to use it
On 26 June, OpenAI previewed GPT-5.6 in three tiers — Sol, Terra and Luna — and described Sol as its most capable model to date, with notable gains in coding, biology and cybersecurity. The headline isn't the benchmark, though; it's the distribution. As CNBC reported, OpenAI is launching only to a "small group of trusted partners" — about twenty organisations whose participation was shared with the government in advance, at the government's request, before general availability "in the coming weeks." This is the first visible invocation of the June executive order's voluntary framework: early government access to frontier models and a vetted-partner release path, with cybersecurity capability as the trigger.
For a European enterprise, the relevant fact is structural, not partisan. Access to the strongest closed model now runs through a list that a foreign government curates and an American company administers, and there is no seat at that table for an EU bank or hospital — only a queue for whenever "general availability" arrives, on terms set elsewhere. It is the same lesson the Fable export-control episode taught in reverse: with a hosted frontier model, the question of whether you can use it at all is decided by someone else's security and trade policy, on their timetable. The everyday case is workable once GA lands; the sovereign case — a regulated workload that cannot wait on, or be re-scoped by, another country's release decisions — is exactly the one this gating leaves unsolved.
Anthropic's Alibaba accusation shows why the walls are going up
The same week, Anthropic accused Alibaba of running the largest model-distillation campaign it has ever documented: operators it links to Alibaba's Qwen lab used roughly 25,000 fraudulent accounts to generate 28.8 million Claude interactions between 22 April and 5 June, routing through commercial proxies to evade the company's China access block. The targeted capabilities were the commercially valuable ones — agentic reasoning, software engineering, long-horizon task completion — and the volume exceeds the combined total of the three distillation cases Anthropic disclosed in February. The allegation went to the US Senate (in a letter to senators) and the White House; Alibaba has not conceded the characterisation.
Read alongside the lead, the two stories are cause and effect. Frontier capability is leaking out of the API faster than vendors can wall it in, which is precisely the argument now used to justify gating who may touch the next model at all. The lesson for a regulated buyer cuts in an awkward direction: the security model of "rent the frontier through an API" assumes the provider can control who is on the other end, and these numbers show how porous that assumption is. A platform whose answer to data control is "trust our access policy" is making a promise that 25,000 proxied accounts just demonstrated is hard to keep — which is why, for the most sensitive workloads, the durable control is hosting the model where the access boundary is your own network, not a vendor's fraud-detection team.
The open-weight escape hatch keeps the frontier reachable
For anyone who isn't on a trusted-partner list, the alternative is the one this brief has tracked all spring: weights nobody can revoke. The standings are now genuinely competitive at the top. MiniMax's open-weight M3, released 1 June, pairs a million-token context window with a reported 59.0% on SWE-Bench Pro — ahead of GPT-5.5 on the company's own numbers — while Z.ai's MIT-licensed GLM-5.2, out mid-June, was independently shown beating GPT-5.5 on SWE-bench Pro at roughly one-sixth the cost. Neither is fresh news this week; the point is the cumulative position, not any single launch.
The structural significance is sharper in light of the lead. When the strongest closed model is available to twenty vetted organisations and the rest of the market waits, a frontier-tier model you can download and run on your own hardware stops being the budget compromise and becomes the only option that is actually under your control. The standing caveats still apply — these are vendor and third-party benchmarks, not your production workload, and the Chinese-origin leaders carry their own provenance scrutiny for EU buyers — so the honest move remains to reproduce the result on your own data before committing a regulated system. But the direction is unambiguous: as access to the closed frontier narrows, the open frontier is the part of the market that no government list can gate.
Quick Hits
- The EU AI Act Omnibus reaches its Council vote on 29 June — after Parliament adopted the text on 16 June, the Council is expected to formally adopt it on Monday, with Official Journal publication to follow. Standalone Annex III high-risk obligations move to 2 December 2027 and the watermarking grace period to 2 December 2026. Two days out, this closes a thread that has run for weeks.
- Gemini 3.5 Pro slips from June to July — Google has quietly pushed general availability into next month, with early testers reportedly flagging token-efficiency and long-horizon task issues. The 2M-token flagship remains in Vertex preview only; the June price-ladder reset many expected now lands in Q3.
- Mistral is reportedly in talks to raise ~€3B at a ~€20B valuation — roughly double its September mark, per TechCrunch and Bloomberg. (Mistral is a European competitor; noted here via third-party coverage, not endorsement.) A well-capitalised EU frontier lab is the supply side of the sovereignty story the rest of this brief is about.
- Claude Fable 5 stays dark — over two weeks after the 12 June export-control order, the model and Mythos 5 remain offline, and Anthropic's US-citizens-only identity-verification path (government ID plus facial geometry) takes effect 8 July ahead of the 1 August executive-order deadline. The standing lesson is unchanged, and the GPT-5.6 gating above is the same lesson from the supply side.