Washington tried to control access to Anthropic's strongest models. The immediate result was not silence, it was a rush of Asian alternatives asking developers to compare price, access, and performance for themselves.
On June 12, 2026, the US government pushed Anthropic's Claude Mythos 5 and Fable 5 behind export controls, and the AI market did what markets usually do when a gate comes down: it looked for another entrance. If you build with these models outside the small circle Washington approves, that shift is not theoretical. Your fallback plan suddenly matters.
Business Insider reported that Anthropic disabled access to Fable 5 and Mythos 5 after a US export-control order barred use by foreign nationals, including some employees inside Anthropic itself. Axios then reported on June 26 that Commerce Secretary Howard Lutnick had approved a limited return of Mythos 5 for named organizations and their foreign-national employees, while Fable 5 remained restricted. That late carveout matters, but it does not erase the bigger message. Access to the best American models can now change because a letter arrives from Washington.
Zhipu AI, the Beijing-based company now branded internationally as Z.ai, moved into that opening with GLM-5.2. The Verge reported on June 28 that Z.ai's open-weight model had narrowed the gap with Mythos in cybersecurity and bug-finding work, even while still trailing Anthropic and OpenAI on broader general tasks. That caveat is important. GLM-5.2 is not magic. But an open-weight Chinese model getting mentioned in the same breath as Mythos on security work is exactly the sort of fact US labs should not wave away.
The price is harder to dismiss. TechRadar reported that GLM-5.2, built on 744 billion parameters and released under an MIT license, had taken the top spot on Design Arena's single-turn HTML web design leaderboard, ahead of Anthropic's Fable 5. It also put the API price at $1.40 per million input tokens and $4.40 per million output tokens, compared with Fable 5 at $10 and $50. You do not need a grand theory of AI sovereignty to understand that math. If the cheaper model is good enough for the workload in front of you, the brand name starts to matter less.
Sakana AI's Fugu is a different answer from Tokyo. The Times of India reported that Sakana launched Fugu on June 22 as a multi-model orchestration system, not a single giant model, with an OpenAI-compatible endpoint that routes work across a pool of models. Its technical report, published on arXiv on June 19, describes Fugu and Fugu Ultra as orchestrator models trained to build agentic scaffolds around each query, pulling in other models where needed. That sounds abstract until you see the use case: one endpoint, multiple models behind it, and less dependence on any single American provider.
On Sakana's own benchmark tables, Fugu Ultra posted 73.7 on SWE-Bench Pro, ahead of Claude Opus 4.8 at 69.2, and the company pitched the system as a way to reach restricted-model performance without using the restricted models. You should treat company benchmarks with care, especially when the company is launching the product being measured. Still, the claim landed because the timing was perfect. Anthropic had just shown the world that model access can be political. Sakana offered a technical workaround with the language of resilience wrapped around it.
Don't overread it. Fugu is still an orchestrator built around models it can reach, and critics have already pointed out that this can replace one kind of dependency with another. The Times of India noted that Fugu Ultra's pricing matched GPT-5.5 at $35 per million tokens combined, so this is not the bargain-bin version of frontier AI. Latency is another real issue. A system that delegates, verifies, and synthesizes across agents can be powerful, but it is not always what you want when a single-model API gives you a fast answer at lower complexity.
Anthropic's own numbers show why this fight is so sharp. The New York Post reported in May that Anthropic had raised $65 billion at a $965 billion valuation, ahead of OpenAI's reported $852 billion valuation, and that its revenue run rate had reached $47 billion. Those are huge figures. They also raise the stakes. When a company at that scale loses global access to its flagship models, even temporarily, every rival gets a line for the sales deck.
Frankly, the export controls gave Asian AI labs the cleanest marketing message they could ask for. Z.ai can say its model is open, cheaper, and strong enough to test. Sakana can say one vendor should not sit between your country and frontier capability. Neither company has proved that it can replace Anthropic across the board, and The Verge's reporting is clear that GLM-5.2 still lags on broader tasks. But the old assumption, that the best model is whatever the leading US lab releases next, looks weaker than it did two weeks ago.
The next test is not another leaderboard. It is whether developers, banks, government agencies, and security teams keep these alternatives in production after the panic fades. If they do, Washington's restriction will have done something unusual: it will have turned a policy meant to protect American AI advantage into a distribution campaign for its rivals.
Also read: Millennium Management is building its own AI lab and that changes the competitive calculus for quant finance; OpenAI just used AI to build its own chip and that changes the quantum threat to crypto faster than anyone planned; BlackRock, Nvidia, and Temasek are betting billions that quantum computing is finally the real thing
