Anthropic launched Fable 5 and Mythos 5 on 9 June. Three days later the US government ordered both pulled from every foreign national worldwide, including Anthropic's own staff. Anthropic had no way to verify nationality in real time, so it pulled both models globally for three weeks.
That's the headline. The part worth a closer look is what sat underneath it.
The order followed a jailbreak report on Fable 5, a prompt technique that bypasses its safeguards. Anthropic tested the same technique against Opus 4.8, GPT-5.5 and Kimi K2.7, and got the same result from all three. This wasn't a Fable 5 flaw, it was a general property of large language models, and it had been known for years.
So the order pulled one company's model for three weeks over a technique the rest of the industry could already reproduce. A jailbreak isn't hardware, it's a prompt, and once it's published everyone has it. The order controlled where Fable 5 could run, not what was already public.
Fable 5 was reinstated for vetted partners on 26 June and the wider controls were lifted on 30 June. The same compute-denial logic behind chip export limits applies here in principle. The problem was never the intent, it was the mismatch between the tool and the target.
Here's who paid for the three weeks, and it wasn't the intended target. Teams that had built workflows around Fable 5 lost the model overnight. The Chinese labs the order targeted kept shipping on schedule.
Nine days after Fable 5 came back, Moonshot released Kimi K3, a 2.8 trillion parameter open-weight model, the largest released to date. It's a sparse mixture-of-experts design, 896 experts, 16 active per token, about 1.8 percent of the network, built for exactly the hardware constraints export controls are meant to enforce.
It scored 88.3 on Terminal-Bench 2.1 against GPT-5.6 Sol's 88.8. It topped LMArena's Frontend Code board. Artificial Analysis placed its composite score within range of Fable 5, a jump of over 700 points from the previous Kimi generation. Moonshot's app already clears $200 million in annualised revenue, backed by Alibaba, Tencent and Meituan, with reported valuation talks near $30 billion. None of that pace came from anything the export control touched.
These are vendor and benchmark-house numbers, self-reported. Take the ranking with that caveat. The scale and the timing aren't in dispute.
Chip restrictions exist to slow exactly this kind of result. Kimi K3 is what it looks like when the slowing doesn't happen.
Fable 5's own return took three deadline extensions before settling on 20 July, into a standard seat on Max and Team Premium plans. Pro and Team Standard get a one-time credit then metered use. Tidy for a subscription tier. Irrelevant to whether the ban's strategy works.
The advice going around is to diversify vendors so no single outage can affect your workflow, which is the wrong problem. Diversification protects against an outage, a rate change, a price rise. It does nothing against a directive that can pull any vendor's model overnight regardless of your alternatives, because the failure wasn't the vendor, it was assuming today's model is tomorrow's model.
The real hedge is architectural, don't build something that depends on one model's exact quirks.
Export controls work on things that stay where you put them. A fabrication plant is fixed and trackable, a demonstrated technique isn't. Once it works against one model it works against the next, and the party left exposed is rarely the intended target. Usually it's the vendor's own customers, waiting on a letter from Washington.
Three weeks bought nobody anything, the capability was already out. The compute the export regime exists to withhold didn't stop a 2.8 trillion parameter model landing four spots off the top either. If the goal was years, Kimi K3 says it's down to months.