The US pulled Anthropic's most powerful model for foreign users — and two open models that can't be revoked

Frontier access just went political. In the same stretch, the US barred foreign users from Anthropic's strongest model, while two open models shipped that nobody can switch off. Here's the builder read on all three.

Here's the 2-minute video version if you want the quick pass first:

1. The US pulled Anthropic's Fable 5 & Mythos 5 for foreign users

A US export-control directive suspended all access to Fable 5 and Mythos 5 by any foreign national — inside or outside the US, including Anthropic's own foreign-national employees.

• With no way to segment foreign users, Anthropic disabled both models for all customers to comply.
• The government cited national security — reportedly a jailbreak method targeting Fable 5's safeguards.
• Anthropic calls it a misunderstanding and is working to restore access. (This is a follow-up to the earlier Mythos / Project Glasswing story.)

Why it matters: Fable 5 had just gone GA days earlier (10 dollars in / 50 dollars out per million tokens, "state-of-the-art on nearly all benchmarks"). A closed, hosted frontier model can be switched off by policy overnight — so it's worth not putting a critical path on a single one.

2. Moonshot shipped Kimi K2.7 Code — a cheap open-weight coder

Moonshot AI released Kimi K2.7 Code on Hugging Face: a coding-first open-weights model anyone can download.

• A 1-trillion-parameter MoE (32B active, 384 experts), 256K context, Modified MIT license.
• Output runs about 4 dollars per million tokens — a fraction of Fable's 50 — undercutting the frontier on price.
• One real caution: it's a Chinese-origin API, so keep sensitive or proprietary code off it.

Why it matters: route bulk work to cheap open models — but vet where a model comes from before you trust it with your code. (Moonshot's benchmark numbers are first-party; no third-party results yet.)

3. Google open-sourced DiffusionGemma — runs free on your own GPU

Google released DiffusionGemma, the first major open-source text-diffusion LLM.

• A 26B A4B MoE that generates 256 tokens in parallel and self-corrects as it goes.
• Over 1,000 tokens/sec on a single H100 — about 4x faster than autoregressive — and small enough to run on a consumer GPU.
• Apache 2.0, natively supported in vLLM, multimodal input.

Why it matters: open weights can't be revoked. Once a model is on your machine, no directive can switch it off — which is exactly the property story #1 makes valuable.

---

Three signals, one pattern: the frontier got pulled, a cheap open model shipped, and a free local one landed. Access to AI is now shaped by geopolitics — so pick a model for **where it runs and who can switch it off, not just its benchmark score. Watch today's full episode, or catch a new one every day on dani / AI News & Creative.