Reuters reported on 30 June that the Commerce Department had lifted export controls on Anthropic's Fable 5 after the lab added safeguards, ending a global shutdown that began with a 12 June national-security directive [E1][E3]. Anthropic said the same day that Fable 5 would return worldwide on paid Claude plans from 1 July, while the more restricted Mythos 5 would widen access to vetted U.S. critical-infrastructure and cyber partners [E2]. The restart matters because Washington had treated a consumer-facing model tier like a controlled dual-use export, rather than ordinary software [E1][E3]. Within days, the same government that flipped access back on was still asking OpenAI to keep its newest frontier preview off the open market [E4].
Meituan, better known as a food-delivery operator than an AI lab, open-sourced a trillion-parameter model trained entirely on domestic Chinese processors without NVIDIA hardware, a release timed to the same policy window [E9]. VentureBeat reported the training run as end-to-end sovereign compute, signalling that frontier-scale work no longer depends exclusively on U.S. silicon supply chains [E9]. A delivery-platform parent building its own chip-cluster training path is structurally different from a pure-play lab racing benchmarks, yet it still advances China's parallel stack [E9]. That juxtaposition is the week's central puzzle: Washington reopened one frontier gate while Beijing demonstrated another training route [E1][E9].
OpenAI's GPT-5.6 Sol preview, announced 26 June, remains limited to a small set of trusted partners whose participation the company said it had shared with the U.S. government [E5]. Reuters reported that broader public rollout was deferred at government request, leaving the most capable OpenAI tier behind an informal trusted-access wall [E4][E5]. Anthropic's stop-and-restart on Fable 5 shows the same pattern in a different costume: globally shippable Sonnet 5 on one side, politically sensitive tiers switched off and on by Commerce letter on the other [E2][E3]. U.S. frontier labs are therefore managing two regulators at once: their own export-control officials and the extraction risk those officials cite [E4][E6].
China's open-weight bench did not need a fresh launch this week to gain leverage; it only needed U.S. access to look conditional. Reuters reported on 25 June that Z.ai's GLM-5.2 had closed much of the frontier gap and framed the release as the first Chinese open model to come that close, with strong coding and cyber performance at a fraction of closed-model cost [E8]. When Anthropic's June shutdown made API availability feel revocable, self-hostable substitutes such as GLM-5.2 immediately became strategic hedges rather than hobby weights [E8]. The practical substitution layer was already in place before Commerce lifted Fable 5 [E8][E2].
Washington's hesitation has documented cause. On 24 June Reuters reported that Anthropic accused Alibaba-linked operators of orchestrating illicit extraction of Claude capabilities, alleging roughly 28.8 million interactions across nearly 25,000 fraudulent accounts; Alibaba denied wrongdoing [E6]. Anthropic had already publicly warned that distillation attacks from Chinese labs threatened the closed-model economics that fund frontier research [E7]. A White House June action on promoting advanced AI innovation and security treats capability access, logging, and nationality verification as national-security infrastructure [E10]. From that vantage, temporary gating buys labs and agencies time to test safeguards, trace misuse, and tighten pre-release review [E10][E4].
Taken together, these facts support a sharper inference than either nationalist cheering or export-control fatalism allows. Chip and model-access curbs were meant to slow China's frontier advance, yet the same curbs make U.S. API dependence feel fragile and reward domestic training plus open-weight redistribution [E9][E8][E3]. Meituan's domestic-chip trillion-parameter release is evidence that export pressure is forging a parallel sovereign compute stack faster than it denies scale, because capital and talent outside traditional labs can now justify frontier training budgets [E9]. Washington's toggle, reopening Fable while holding GPT-5.6 in trusted preview, trains every non-U.S. buyer to hedge with local silicon and downloadable weights [E1][E4][E8].
The counter-case deserves equal weight. Commerce did not lift Fable blindly: Reuters tied the restart to added Anthropic safeguards, implying that oversight worked rather than collapsed [E1][E2]. A trillion parameters trained without NVIDIA proves domestic scale while still falling short of automatic parity with the closed frontier; Reuters framed GLM-5.2 as unusually close, still below equivalence to Anthropic's sensitive tiers [E8][E9]. Distillation at the scale Anthropic alleged, if validated, would confirm why governments treat API access as a contested extraction surface worth pausing [E6][E7]. Export controls may accelerate China's parallel stack, yet they still purchase security review time that a week of headlines cannot dismiss [E10][E1].