In the two weeks since the US government restricted global access to Anthropic's Mythos and Fable 5, a Japanese startup and a Chinese lab have each released models that match or beat them on benchmark after benchmark, at prices that make the American offering look extravagant.
On June 12, 2026, the US Department of Commerce placed export controls on Anthropic's Mythos and Fable 5 models. Teams outside a narrow list of vetted US organizations lost API access overnight. What the restrictions exposed, almost immediately, is how thin the moat actually was.
Zhipu AI, the Beijing-based lab that operates under the brand Z.ai, released GLM-5.2 on June 13, the day after the ban. It's a 744-billion-parameter mixture-of-experts model, roughly 40 billion active parameters per token, with a one-million-token context window and an MIT license. You can download the weights from Hugging Face for nothing. On SWE-bench Pro, the standard long-horizon software engineering benchmark, GLM-5.2 scores 62.1, ahead of GPT-5.5's 58.6. On AIME 2026, the hardest public math reasoning test, it hits 99.2, above GPT-5.5's 98.3. The API price is roughly one-sixth that of GPT-5.5. As VentureBeat noted, Semgrep's own internal cyber benchmarks put GLM-5.2 ahead of Claude on several measures, prompting the security research team to publish a post titled, bluntly, "We have Mythos at home."
Then, nine days later, Tokyo-based Sakana AI released Fugu. Where GLM-5.2 is a single scaled model, Fugu is something architecturally different: a multi-agent orchestration system that coordinates a pool of existing frontier models through components the company calls TRINITY and Conductor, presenting a single OpenAI-compatible endpoint to the developer. The marketing copy on Sakana's site advertises "frontier capability without the risk of export controls," which is blunt even by startup standards. Across 11 major benchmarks, Fugu and its heavier sibling Fugu Ultra post top scores on 10 of them, including a 73.7% on SWE-Bench Pro, ahead of Claude Opus 4.8's 69.2% and well ahead of Gemini 3.1 Pro's 54.2%. On the benchmark tables Sakana published, Fugu Ultra matches Anthropic's Fable 5 and Mythos Preview, the two models Fugu's users in Japan and Southeast Asia can no longer access.
Anthropic's situation is worth being direct about. The company recently closed a funding round at a $965 billion valuation, ahead of OpenAI's $852 billion. Its revenue run rate reached $47 billion. These are extraordinary numbers. But for developers and enterprises outside the Commerce Secretary's approved list, the practical effect of the June 12 restrictions is simple: the best models Anthropic has built are off the table. That is a market opening, and Sakana and Zhipu walked straight through it.
A Sakana spokesperson told reporters last week that "US models remain important to Asia," which is the diplomatic version of saying the door hasn't closed permanently. Co-founder Ren Ito made similar remarks at the G7 summit in Evian, where export controls and AI access were a central topic. Don't mistake the diplomatic tone for a lack of ambition. Fugu launched with explicit language about AI sovereignty, and Sakana is clearly targeting Japanese government agencies and enterprises that need a frontier-class model they can actually use and not worry about losing to a Commerce Department decision next quarter.
The cost gap is the part that matters most in the long run. GLM-5.2 at one-sixth the price of GPT-5.5, with benchmark performance that meets or exceeds it on the workloads most enterprises actually care about, is not a marginal alternative. That's a real substitution. And because Zhipu released the weights under an MIT license, any company with the infrastructure can run it without paying an API fee at all. Anthropic charges $25 per million input tokens for Mythos Preview and $10 per million for Fable 5. The math isn't close.
Worth keeping honest: Sakana's benchmark claims met skepticism quickly. Within 24 hours of Fugu's launch, independent testers reported a gap between the published numbers and real-world use, a recurring problem in the industry that applies to US labs as much as Japanese ones. Real-world coding performance and benchmark scores are not the same thing, and Fugu's multi-agent architecture introduces latency that a single-model API does not. GLM-5.2 has its own caveats, including questions about how reliably the 1M-token context window performs at full length. But the fundamental story isn't changed by those caveats. Two weeks ago, Mythos and Fable 5 were the only frontier-class models with these capability profiles. Today there are credible alternatives that most of the world can actually access.
Frankly, the export controls may have handed Asian AI labs the best marketing they could have asked for. The restrictions created an urgent, concrete reason for developers in Japan, Southeast Asia, and Europe to evaluate alternatives they might otherwise have treated as distant second choices. Sakana's and Zhipu's timing, whether calculated or fortunate, was near-perfect. US labs now face pressure from two directions at once: match the cost efficiency, and do it without hiding behind access restrictions that competitors are explicitly positioning against. The next few months will show whether that pressure produces real price movement or just a quieter version of the same market structure.
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
