Sign in Subscribe
AI Intelligence

U.S. Model Controls Open a Door China and Japan Walk Through - Week of June 21 - June 27, 2026

Vijay Bhagwati

• 13 min read
Share

Week of June 21 – June 27, 2026

Reading time: 12 minutes

The most important shift this week is not a model release or a funding round. It is the return of export controls as the central market force in artificial intelligence. Two weeks after the Trump administration forced Anthropic to pull its Fable 5 and Mythos 5 models from non-U.S. users, Washington asked OpenAI to throttle the rollout of GPT-5.6 as well. The stated reason is national security: frontier models strong enough to assist offensive cyber operations or biological weapon design should not move freely. The practical result is different. By limiting who can use the most capable closed U.S. models, the controls are creating a demand vacuum that open-source and semi-open alternatives from China and Japan are filling at speed. Intelligence per dollar, not raw capability, is becoming the decisive buying criterion. That reorders incentives for labs, enterprises, chipmakers, and governments.

Anthropic began the week under pressure. On June 24 it sent a confidential letter to Senators Tim Scott and Elizabeth Warren alleging that operators affiliated with Alibaba and Alibaba Qwen used almost 25,000 fraudulent accounts to conduct more than 28.8 million exchanges with Claude between April 22 and June 5, in what Anthropic described as the largest capability-extraction campaign it had ever measured (Ars Technica, June 25). The target was not general chat but "agentic reasoning, software engineering, and long-horizon tasks." Anthropic framed the operation as circumvention of U.S. access restrictions and called for punitive action against Alibaba. On June 26 the Commerce Department partially relented, allowing Anthropic to restore Mythos 5 access to "a limited number of trusted users," and Axios reported on June 27 that Fable 5 limits could be lifted as early as the coming week. The thaw is narrow: trusted users only, with guardrails. It is also fragile, because the episode revealed that U.S. frontier labs now face adversarial distillation at industrial scale.

OpenAI followed a similar arc. On June 26 it previewed GPT-5.6 in three tiers — Sol, Terra, and Luna — and simultaneously disclosed that the U.S. government had asked it to limit initial access to roughly 20 pre-approved companies (Axios, June 26; OpenAI blog). CEO Sam Altman had previewed the model with the White House in early June. The goal is still a broad release, but Washington's review process has inserted itself between product completion and commercial availability. OpenAI also suffered a personnel leak: Bloomberg reported on June 26 that Paul Meade, Apple's Vision Pro and smart glasses chief, is leaving for OpenAI, while TechCrunch reported the same day that OpenAI hired Prabhjeet Singh, Uber's India and South Asia president, as its first managing director for India. The hires signal hardware ambition and a bet on India as the largest non-U.S. market.

The week also delivered the most credible open-source challenge to U.S. frontier models since DeepSeek V4. Zhipu, the Chinese startup backed by Alibaba and Tencent, released GLM 5.2 under an open license. CNBC reported on June 26 that the model sits within one percentage point of Anthropic's Opus 4.8 on a key agentic benchmark at roughly one-fifth the cost, with OpenRouter token traffic climbing faster than after DeepSeek's April launch. Zhipu's timing is not accidental. With Anthropic's Fable and Mythos restricted and OpenAI's GPT-5.6 gated, a downloadable model that cannot be revoked looks safer to enterprises worried about access continuity. Harvey co-founder Gabe Pereyra told CNBC, "GLM 5.2, you're seeing the first model where it's really competitive with some of these closed-source frontier models." The phrase "intelligence per dollar" appeared repeatedly in coverage, suggesting the buyer's metric has shifted.

Japan joined the race from a different angle. Sakana AI, the Tokyo-based lab known for evolutionary model merging, launched Fugu Ultra on June 22 and positioned it as a multi-agent orchestration model that can match Fable 5 and Mythos Preview on engineering and scientific benchmarks without single-vendor dependency (Sakana AI blog). Fugu is not a raw foundation model in the traditional sense; it is trained to dynamically call and orchestrate other models, including recursive instances of itself. The framing is deliberate: U.S. export restrictions become irrelevant if customers can route to a constellation of models through a single API. DeepSeek, meanwhile, open-sourced DeepSpec, a full-stack training and evaluation framework for speculative decoding draft models, including the DSpark algorithm that Hacker News upvoted to the top of the week with 755 points. Speculative decoding is an inference-efficiency play, exactly the kind of technology that makes open models cheaper to run at scale.

Hardware and physical AI had their own week. ON Semiconductor agreed to buy Synaptics for roughly $7 billion in a deal explicitly framed as a push into physical AI — sensors, touch, and edge compute for robots, cars, and devices (CNBC, June 25). The transaction is a bet that AI value will migrate from training clusters to the physical layer: cameras, haptics, motor controllers, and the chips that bind them. Google added computer-use capabilities to Gemini 3.5 Flash, enabling the model to control a desktop environment (Google blog, June 24), though Business Insider reported the same week that Gemini 3.5 Pro has slipped to July. Micron overtook Meta in market value, underscoring memory's centrality to the AI build-out. And Oracle ended its worst week since the 2001 dot-com bust as investors worried about data-center financing, a reminder that the infrastructure boom is not costless.

The policy layer thickened. California launched a state-level AI job-loss tracker as layoff fears spread (Bloomberg, June 25). The Pentagon is reportedly seeking a broader role for AI in target identification (Bloomberg, June 25). Politico reported on June 27 that Silicon Valley is struggling to navigate the Trump administration's AI about-faces. Europe watched the U.S.-China contest from the sidelines: Mistral CEO Arthur Mensur argued that AI companies should pay a content levy in Europe, while data-center lobbyists warned that Europe must choose between AI and climate goals. The EU is still regulating; it is not yet competing at the frontier.

Ten Pillar Review

Frontier Models. The dominant fact is fragmentation. U.S. labs remain at the cutting edge on raw benchmarks, but their ability to monetize that edge is now conditional on federal approval. OpenAI's GPT-5.6 Sol is the strongest announced model of the week, yet it is available only to a government-curated shortlist. Anthropic's Mythos 5 is partially restored to trusted users. Meanwhile Zhipu's GLM 5.2 and Sakana's Fugu Ultra are available globally. The gap between frontier and near-frontier is narrowing, and access has become a product feature.

Open Source. This was the strongest open-source week of the year so far. Zhipu GLM 5.2 is the headline, but DeepSeek DeepSpec and Sakana Fugu show the movement diversifying. Open-source pressure is moving from chatbot performance to agentic capability, inference efficiency, and model orchestration — the exact capabilities U.S. export controls are trying to bottle up. The Chinese and Japanese alternatives are not merely cheaper; they are positioned as more durable against political interruption.

Agentic AI. OpenAI published a 50-page paper, "The Shift to Agentic AI: Evidence from Codex," finding that active Codex users grew more than fivefold in the first half of 2026, that more than 10% of users manage three or more concurrent agents, and that 26.6% use "skills" for complex workflows (OpenAI research PDF, June 25). Within OpenAI, 97.9% of employees now use agents and they have "largely replaced business usage of ChatGPT." Google added computer use to Gemini 3.5 Flash. Patronus AI raised $50 million to build simulated digital worlds for stress-testing agents (TechCrunch, June 25). Microsoft gave 33-year-old executive Jacob Andreou responsibility for fixing Copilot, with Satya Nadella reportedly impressed by the speed at which his team shipped the agentic Copilot Tasks feature (Fortune, June 27). Agents are no longer a demosphere; they are becoming the primary interface for knowledge work inside the most AI-forward organizations.

Frameworks. DeepSpec is the standout. It provides data preparation, training, and evaluation scripts for speculative decoding draft models, with a reported target-cache footprint of roughly 38 TB for the default Qwen3-4B setting. The release signals that the open ecosystem is investing in the boring but critical layer of inference infrastructure, not just headline models. WorkWeave's router, a Show HN project that gained 203 upvotes, adds smart model routing inside Claude, Codex, and Cursor. Enki, another Show HN project, benchmarks memory compression for agents. The tooling layer is maturing around efficiency and routing.

Hardware. ON Semiconductor's $7 billion Synaptics bid is the week's largest hardware story and a clear statement that edge AI and physical sensing are the next growth pools. SK Hynix is targeting a $29 billion U.S. listing. Micron's market-cap ascent past Meta shows how memory constraints are reshaping the semiconductor pecking order. NVIDIA remained mostly absent from the headline cycle, though a Tom's Hardware report that five-year-old A100 servers now sell for up to $82,000 in China illustrates the scarcity premium still attached to U.S. export-controlled silicon.

Economics. Two competing forces are visible. On one side, SoftBank shares fell after Bloomberg reported an OpenAI IPO delay. Oracle had its worst week since 2001. AI infrastructure financing is under scrutiny. On the other side, Redo, a Utah e-commerce technology company, raised $81 million at a $1.25 billion valuation to build an AI-powered customer-retention platform, and Patronus AI's $50 million round suggests robust demand for agent evaluation. The capital market is differentiating: infrastructure spending is being questioned; application-layer revenue is still being rewarded.

Physical AI. ON Semiconductor-Synaptics, Gemini 3.5 Flash computer use, and the arXiv paper "E-TTS: A New Embodied Test-Time Scaling Framework for Robotic Manipulation" (2606.27268) all point in the same direction. Value is moving from pure cognition toward systems that can perceive and act in the physical world. Robotics remains harder and more capital intensive than software agents, but the week showed progress in both the commercial layer (edge sensors, computer use) and the research layer (embodied test-time scaling).

Security. Anthropic's Alibaba allegation is the week's sharpest security story: capability extraction at scale through fraudulent accounts. It demonstrates that model security is no longer just about weights and endpoints; it is about adversarial usage of legitimate APIs for distillation. The Commerce Department's partial restoration of Mythos 5 to trusted users suggests the U.S. is trying to balance security with commercial viability, but the difficulty of that balance is now exposed.

Sovereign AI. India is the clearest case. OpenAI's hire of Prabhjeet Singh, plus its existing New Delhi office and planned Mumbai and Bengaluru offices, treats India as a strategic non-U.S. market. China's Zhipu and 360's Tulongfeng are sovereignty plays in the opposite direction: domestic alternatives to U.S. frontier models made necessary by U.S. restrictions. Japan's Sakana Fugu is a sovereignty play of a third kind: a U.S.-allied democratic country building frontier alternatives through model orchestration rather than scale. The world is splitting into AI spheres faster than any single regulatory framework can manage.

Enterprise AI. The enterprise story is cost pressure plus access risk. CNBC's Zhipu coverage cited enterprises hit by unexpectedly high AI token spend who are now asking how to maximize intelligence per dollar. Salesforce employees are reportedly concerned about Anthropic's expansion into Slack (The Information, June 27). California's job-loss tracker reflects a political system beginning to count the labor displacement cost. Enterprises are still adopting AI, but they are doing so with sharper procurement discipline and greater concern about supplier lock-in and revocation risk.

Pattern Shifts

Accelerating. Adversarial distillation, agentic adoption inside tech companies, open-source model quality in China and Japan, inference-efficiency tooling, physical AI M&A, and sovereign AI investment. The most underreported acceleration is the normalization of government review as a release gate. Whether the policy is wise, it is now a structural feature of the U.S. AI product cycle.

Stalling. U.S. frontier model monetization, at least internationally. Anthropic and OpenAI are spending weeks negotiating access rather than selling it. Google Gemini 3.5 Pro has slipped to July. Oracle's financing concerns suggest data-center build-outs may face a capital-market pause. The pure "scale at all costs" narrative is encountering both political and financial friction.

Surprises. That Zhipu, not another DeepSeek release, would produce the strongest open-source challenge to Anthropic and OpenAI. That Japan's Sakana AI would frame its answer to U.S. export restrictions as a multi-agent orchestration model rather than a bigger monolithic model. That Microsoft would turn to a 33-year-old executive to rescue Copilot, a product on which the company's AI strategy depends. That California, not the federal government, would build the first statewide AI job-loss tracker.

Contrarian signals. The Hacker News audience upvoted DeepSpec (speculative decoding) and WorkWeave (model routing) more than it upvoted the GPT-5.6 announcement. That is a developer-base signal that efficiency and control matter more than marginal benchmark gains. Meanwhile, Anthropic's own research showed that agentic usage is growing fastest outside the original developer audience, which means the next wave of demand may come from non-technical workflows, not coding.

Breakthrough Papers

"Reinforcement Learning without Ground-Truth Solutions can Improve LLMs" (arXiv:2606.27369). Yingyu Lin, Qiyue Gao, Nikki Lijing Kuang, Xunpeng Huang, Kun Zhou, and colleagues introduce RiVER, a ranking-induced verifiable-reward framework that trains LLMs on score-based optimization tasks without ground-truth solutions. On 12 AtCoder Heuristic Contest tasks, RiVER improves Qwen3-8B and GLM-Z1-9B-0414 by 8.9% and 9.4% in ALE rating rank, and it transfers gains to exact-solution benchmarks such as LiveCodeBench and USACO. The implication is that RLVR can work in settings where correctness is not binary, widening the set of tasks that can be used for post-training.

"E-TTS: A New Embodied Test-Time Scaling Framework for Robotic Manipulation" (arXiv:2606.27268). The authors propose a reasoning mechanism for embodied tasks that incorporates historical information and test-time scaling. The advance is the recognition that embodied reasoning is not just about planning the next action but about maintaining and querying a history of interaction, with a scaling law for how much reasoning improves performance. It is one of several papers this week pointing toward test-time scaling as the next lever beyond pre-training scale.

"Hallucination in World Models is Predictable and Preventable" (arXiv:2606.27326). The paper argues that generative world models hallucinate in low-coverage regions of the state-action space and that lightweight density estimation can predict and prevent these failures. As world models become central to robotics, simulation, and planning, understanding where they lie becomes as important as improving average-case accuracy.

"Paved with True Intents: Intent-Aware Training Improves LLM Safety Classification Across Training Regimes" (arXiv:2606.27210). The authors introduce AIMS, a dataset of 1,724 difficult safety prompts annotated with user intent, and show that modeling intent as an explicit signal improves safety classification across supervised, few-shot, and reinforcement-learning training regimes. The work matters because current safety systems often fail on indirect or coded language; intent-aware training is a direct response.

Falsifiable Predictions

  1. Anthropic will restore broad Fable 5 access, not merely trusted-user access, within 14 days of June 28. The Axios report and Commerce Department partial reversal point to a near-term thaw, and the revenue and customer lock-in cost of keeping the model offline is large.
  2. OpenAI will broaden GPT-5.6 access beyond the initial 20 approved companies before July 15. OpenAI's own blog said it expects to expand access to more companies next week and that the government has expressed support barring concerns in additional testing.
  3. Zhipu GLM 5.2 or a successor will crack the top three on at least one major agentic benchmark within 60 days. The CNBC report already places it within one percentage point of Opus 4.8; the trajectory and OpenRouter traffic growth support continued improvement.

Sources

Published June 28, 2026. This briefing is for informational purposes and does not constitute investment, legal, or policy advice.