Week of 2026-06-15 to 2026-06-21 · The US kept Anthropic’s two best models dark all week. The demand didn’t wait — it walked to open weights, model-agnostic harnesses, and local rigs.

Last week the US government switched off a commercial AI model for the first time. This week we found out what happens next.

Quick recap. On June 12, Commerce invoked national-security authority and ordered Anthropic to suspend Fable 5 and Mythos 5 for every foreign national — including its own foreign-born staff. Unable to geofence its employees out of its own models, Anthropic disabled both globally. The trigger was a reported jailbreak of Mythos’s cybersecurity guardrails. Anthropic called it narrow and warned that if one narrow bypass justifies recalling a model served to hundreds of millions, the same standard “would essentially halt all new model deployments.” We covered the politics last week. This week was about the consequences.

They were not the consequences the ban was for.

The demand didn’t disappear. It relocated.

Four days after the most capable closed model Americans could all legally use went dark, a near-frontier replacement shipped — for free.

On June 16, Z.ai released GLM-5.2 under an MIT license: full open weights, a 744B-parameter mixture-of-experts model with ~40B active per token and a 1M-token context window. Artificial Analysis ranks it the top open-weight model on its Intelligence Index — 51 on v4.1, an 11-point jump over GLM-5.1 — and effectively level with GPT-5.5 on its real-world agentic benchmark (GDPval-AA: 1524 to GPT-5.5’s 1514). It sits on the intelligence-versus-cost frontier. Simon Willison called it “probably the most powerful text-only open weights LLM.” You can download it. No one can switch it off.

Practitioners read the contrast immediately. “Has anyone replaced Claude/GPT with a local model for daily coding?” hit 540 points and 277 comments on June 16. A month ago that’s a hobbyist question. This week it was the room checking the exits.

And the exit they reached for most isn’t a model — it’s a harness. OpenCode, the model-agnostic, MIT-licensed terminal agent, passed Claude Code on GitHub stars: roughly 172,000 to Claude Code’s ~124,000. Stars are a vanity metric — usage and stars are not the same number — but the direction is the signal. The tool developers are bookmarking is the one that doesn’t care which model you run. OpenCode spent this quarter on the receiving end of Anthropic’s OAuth blocks and legal demands; people are buying option-value against their own vendor, and now against their vendor’s government too.

The pattern underneath all three: a trained model is non-excludable once it’s downloadable. Switch off the legible artifact and users route to a substitute — an open model, a portable harness, a local rig. The ban contained exactly one thing: Anthropic’s own addressable market.

This is the channel thread, confirmed live

We’ve argued for a month that the moat is the channel, not the weights. Last week the political off-ramp opened — controls hit the closed US leader while open Chinese weights walked free. This week we watched users actually take it. The state tried to make the trained artifact the chokepoint, and the artifact turned out to be a pile of numbers you can mirror. That isn’t a metaphor; it’s the crypto-wars precedent we mapped two weeks ago — control the math, and the math leaves the country anyway.

The uncomfortable part for Anthropic: it is right on the merits and losing on the board. The jailbreak standard is unworkable. The ban is bad policy. And being the responsible, legible, US-based lab is now a structural tax — the more switchable-off you are, the less you are worth as a dependency. Safety was the brand. This quarter it became the liability, and this week the bill arrived in the form of a competitor’s MIT-licensed download counter.

The squeeze came from the other side too

The supply shock landed in the same seven days the demand-side economics tightened.

OpenAI’s leaked financials — reported by Fortune and Ars Technica, not independently verified, and disputed in the details — show roughly a $21B operating loss on ~$13B of 2025 revenue. Net-loss figures range wildly depending on how you treat valuation accounting, but the shape is not in dispute: the frontier is sold well below cost. The same week, the FT reported enterprises are reining in AI spend as bills strain budgets. And Anthropic’s own subscription split took effect June 15, moving programmatic usage into a separate credit pool — repricing exactly the agent-heavy workflows that flat-rate plans had been quietly subsidizing.

Put it together. The legible US leader took a supply shock — its two best models dark — while its customers took a cost shock — metering tightening, budgets pulling back — in the same week a free, downloadable, near-frontier substitute shipped under the most permissive license there is. If you set out to design a week that accelerates migration off closed frontier APIs, this is the week you’d draw up.

What the reader should do

Provider-portability stopped being hygiene and became risk management.

The lesson is not “switch to open weights.” GLM-5.2 is brutal to run locally at 744B parameters — its sparsity makes it cheap to serve at batch scale and punishing to host on your own hardware — and closed frontier models still win most real work. The lesson is narrower and sharper: your best model can now be switched off by someone who is neither your vendor nor your government. The trigger here was a tip from Amazon — Anthropic’s largest investor and a model competitor. Design for that.

A multi-provider gateway and a model-agnostic harness are now how you price in political risk, not just latency and cost. Concretely: keep your prompts and tools provider-portable, keep one open-weight fallback wired up even if you never run it in anger, and treat “what happens to this workflow if this specific model goes dark on a Friday” as a real line in your design review. It happened to the best-resourced lab in the country this month. It can happen to your dependency.

Also this week

  • Wired named the trigger. SK Telecom was reportedly the company whose Mythos jailbreak demonstration set the Commerce order in motion; SK denies China ties. Worth noting how thin the public chain from “one demo” to “switch off a model for everyone” still is.
  • The regime can’t aim. Commerce held off blacklisting DeepSeek specifically while adding 100+ firms to the entity list. The order that darkened a closed US model can’t bring itself to directly hit the open Chinese one — because there’s no artifact to seize.
  • Brain drain accelerates. Noam Shazeer joined OpenAI as Lead for AI Architecture Research — less than two years after Google paid ~$2.7B to bring him back. The co-author of both founding MoE papers leaves the week GLM-5.2’s MoE design tops the open-weight charts. Fitting.
  • A real Git challenger, from a real shop. Epic open-sourced Lore, an MIT-licensed Rust VCS for large binaries (859 points on HN). Game-scale by design, general-purpose in practice — the first serious Git alternative backed by a studio with the assets to prove it.
  • The anti-hype data point. DuckDuckGo’s Gabriel Weinberg argued AI adoption is segmented, not universal — the hottest HN thread of the week (391 points). A useful counterweight in a week where the news read like the whole economy runs on tokens. It doesn’t, yet.

Ledger note: nothing came due this week. The W23 Copilot-walkback call is due ~Jul 5 (still open, no reversal yet); the W24 export-ban-narrowing call is due ~Aug 14. Both settle in a later issue. Scorecard stays 0–0.

One thing to watch

The ban’s lasting market effect won’t be a capability dent — the capability is downloadable. It’ll be a portability premium.

Prediction (60% confident): within 90 days (by ~2026-09-20), at least one major commercial AI vendor (Anthropic, OpenAI, Google, or Microsoft) ships or formally announces a customer-facing multi-provider / bring-your-own-model fallback in a first-party developer product — positioning provider-portability as resilience against exactly the switch-off risk this week made concrete. If no such move appears and first-party tools stay locked to a single house model, I’m wrong.