The US Banned Fable 5. Open-Weight Models From China and Canada Took Over in 6 Days.

by Persephone

Six days into the Fable 5 export ban, enterprise engineering teams are running on Cohere, Moonshot, Zhipu, and Llama 3 — and Anthropic is in DC negotiating its way back in.

The export-control regime was supposed to keep frontier AI inside US borders. Six days in, it has done the opposite: pushed enterprise AI workloads onto open-weight models from Toronto, Beijing, Hangzhou, and Menlo Park. On June 12, 2026, the US Commerce Department under Howard Lutnick disabled Fable 5 and Mythos 5 for foreign-national access at 5:21 PM Eastern. By Day 6, the four open-weight coding models — Cohere’s Command, Moonshot’s Kimi, Zhipu’s GLM, and Meta’s Llama 3 — had absorbed the displaced production traffic. The ban created the exact outcome it was meant to prevent.

The timeline, fast

  • June 12, 5:21 PM ET — Commerce Secretary Howard Lutnick issues an emergency export-control directive. Anthropic disables Fable 5 and Mythos 5 globally for any user matching the foreign-national criteria. The government’s stated trigger: a jailbreak path it claims to have identified, attributed to a South Korea-based origin.
  • June 13 — TopClanker’s first post on the ban frames it as a “deemed export” problem: a US-built model, served from US infrastructure, restricted from US persons working abroad.
  • June 15 — The follow-up notes that the Fable 5 jailbreak is functionally the product. The “fix” the government demanded was the same capability the model was designed to provide to approved defenders under Project Glasswing.
  • June 17, Seoul — Anthropic opens a Seoul office. International Business GM Chris Ciauri tells Korean press the company is “very confident these models will be available again within the next few days.”
  • June 18, today — Anthropic teams are in Washington for face-to-face talks with Commerce and the NSC. The next 48 to 72 hours decide whether Fable 5 comes back, comes back constrained, or stays off.

This is the saga the blog has been tracking. The story has moved from “Anthropic got caught” to “what enterprise teams actually did in response.”

What enterprise teams actually did

They did not wait.

Per forgenex.com reporting, Llama 3 and Mistral were the immediate fallbacks within 24 hours — weights downloadable, self-hostable, no export-control surface. By Day 3, engineering leads began the second-source work that most of them had been deferring for a year: routing production traffic through Cohere’s Command, Moonshot’s Kimi K2, and Zhipu’s GLM-5. These are not hobbyist models. Cohere is headquartered in Toronto. Moonshot and Zhipu are Chinese labs with documented frontier-class coding benchmarks. The traffic that used to terminate on Anthropic’s API now terminates on open-weight weights running on the buyer’s own infrastructure, or on a non-US provider that the US government cannot reach with a single directive.

The New Stack’s coverage gets the framing right: four open models responded before Anthropic could restore access. The interesting detail is not the number. It is the speed. Enterprise teams had a tested second-source plan ready because they had been quietly building one since the first Model Spec dispute. The ban did not create the alternative. The ban removed the friction that was suppressing the alternative.

Project Glasswing: US defenders cut off from their own tool

The sharpest cut is on the US side. Project Glasswing is Anthropic’s controlled-access program that gives cyber-defense teams and approved researchers a Mythos-class model with the safety guardrails tuned for offense-side work. Fable 5 is the same underlying model as Mythos 5, packaged for broader use with safeguards. The government’s directive pulled both.

Luta Security, which covers the defender community closely, ran the numbers: the ban cuts US cyber defenders off from a tool they were already running in their stack, while the same capability remains available in GPT-5.5, Gemini 2.5 Ultra, and any sufficiently motivated open-weight deployment. Per explainx.ai, Anthropic’s defense in the Commerce talks is straightforward — the export control is non-unique. The capability the directive was meant to restrict is already on every commercial frontier API and is reproducible from public model weights. The directive does not constrain the adversary. It constrains the defender.

This is the line Anthropic is taking to Lutnick’s office this week.

The open-weight political play

There is a second-order story that matters more than the saga itself. While the G7 debates export-control harmonization, China’s open-weight labs are pitching free AI infrastructure to the developing world. Moonshot and Zhipu are the most visible, but the pattern is broader: open weights become a soft-power tool the moment the US restricts its own frontier APIs. The result of the Fable 5 ban, six days in, is a US export-control regime that has measurably accelerated the adoption of Chinese open-weight models by enterprise teams in friendly jurisdictions.

That is not what the directive was meant to do.

What to do this week

If you are a platform engineering lead running Fable 5 in production, you have probably already moved. If you have not, here is the practical sequence:

  1. Make the second-source permanent, not provisional. Cohere Command, Moonshot Kimi, Zhipu GLM, and Llama 3 are now the actual production stack. Treat the Fable 5 path as a degraded fallback until Anthropic publishes a written re-enablement notice, not a verbal one.
  2. Audit your export-control exposure. If any of your inference paths route through a US provider that may receive a Lutnick-style directive, the answer is to push the weights to your own infrastructure, not to negotiate a letter of assurance from a vendor.
  3. Pressure-test your cyber-defense tooling. Project Glasswing customers in particular need to know whether the guardrailed model they were paying for is the same model now on a US government blocklist. If you are defending US critical infrastructure, you are now in a worse position than you were on June 11. That is the part the directive is not acknowledging.
  4. Plan for a constrained re-enablement. The most likely outcome of the DC talks is Fable 5 returns with tighter foreign-national gating, not a clean restoration. Build your 30-day plan around that, not around Ciauri’s “next few days.”

The export-control regime was sold as a way to keep American AI capability inside American control. Six days in, it has transferred that capability to open-weight providers in Toronto, Hangzhou, and Beijing. The next two to three days will tell us whether the US government has noticed.

Sources