1. China's answer ships as a free download

Forty-eight hours after Washington pulled Anthropic's best model, China gave its best one away. Z.ai released GLM-5.2, an open-weight model (MIT license) ranked the most capable openly available and priced at roughly a tenth of Claude or GPT-5. If the model you depend on can be switched off, a frontier-grade one you can host yourself stops looking like a hobby and starts looking like a hedge.

  • The trade-off is jurisdiction, not quality: self-host and the data stays yours; use Z.ai's API and prompts route through servers under China's intelligence law, with Z.ai on the US Entity List. That belongs in diligence, not marketing.

2. Fable's still dark, the precedent isn't

Fable 5 and Mythos 5 remain suspended worldwide under the US export-control directive; Anthropic says access could return within days, and markets price roughly 57% before July 1. Treat it as unavailable until confirmed. A "Fable is back" report this week did not hold up. The durable lesson is the precedent, not the outage: a frontier model can be ordered offline overnight, and "who decides what ships" is now a planning assumption.

  • A quiet escalation: Anthropic's updated terms (effective July 8) add government-ID and biometric collection for verification.

3. AI bosses tell the G7: govern us

At the G7 in Evian on June 17, the three rival lab chiefs told heads of state that frontier systems are too consequential to be governed by their builders alone, and Anthropic's CEO called for a US-led standards coalition five days after the US forced his own models offline. The tension is the story: the labs are inviting oversight in principle the same week they resist it in practice.

  • For anyone building on these models, the rules of access are now set in capitals, not just product roadmaps.

4. SpaceX buys Cursor for $60B

Days after its record IPO, SpaceX (xAI's parent) agreed to acquire Cursor maker Anysphere in an all-stock deal valuing it at $60B, expected to close in Q3. The race is now being fought over the harness, the tool wrapped around the model, as much as the model itself: SpaceX now owns its compute, the most-used enterprise coding harness, and the distribution for both.

  • The control point keeps moving up the stack: cheaper inference and the interface developers live in may matter more than the single best raw model.

5. The model flood is loaded

The capability race runs on regardless of who is banning or buying whom: GPT-5.6 is tracking a June 22-28 launch at reportedly a third of Fable's price, Gemini 3.5 Pro is due for general availability this month, and Google's Antigravity CLI just replaced the Gemini CLI. Three of the year's strongest models may land in a single week.

  • The question still stands: are your workflows model-agnostic enough to swap on a step-change, or to switch providers if one gets pulled?

More AI stories of the week

  • AMD blows past a $900B valuation. A record high on its MEXT memory-software acquisition and the launch of the $3,999 Ryzen AI Halo, a desktop box that runs 200-billion-parameter models locally and undercuts Nvidia's equivalent by $700.
  • Colorado's AI Act takes effect June 30. The first broad US state law on high-risk AI (hiring, lending, insurance, healthcare) goes live in ten days, even as Washington's effort to preempt state AI laws remains unresolved; compliance teams cannot wait for that fight to settle.
  • SAP acquires Prior Labs. The enterprise-software giant is buying the German tabular-data lab and pledging over $1.18B to build an in-house frontier capability specialized in structured business data, a sign incumbents want to own the model, not just rent it.
  • Meta locks in 6 gigawatts of AMD GPUs. A multi-year MI450 deal (plus a separate ~$27B Nebius compute pact) shows the buildout is now sized like national infrastructure, with chip supply locked years ahead.
  • Local inference goes mainstream. Between AMD's Ryzen AI Halo and Nvidia's DGX Spark, a ~$4,000 machine can now run a 200B-parameter model on your desk, turning "run it yourself" from a sovereignty slogan into a purchase order.
  • DeepMind maps the road past AGI. A new research report lays out four pathways from human-level AGI to artificial superintelligence; less a product than a signal of where the labs think the frontier is heading.
  • Europe 2031. A "future history" manifesto arguing Europe risks AI irrelevance despite its talent and data, and what it would take to change course; useful framing ahead of the Paris forum.
  • Midjourney launches a full-body medical scanner. The image-AI lab is investing $74M+ in "ultrasonic CT," a radiation-free 60-second whole-body scan it claims rivals MRI, with a first clinic in San Francisco; a striking bet that AI labs will move into physical hardware.
  • Google and Kaggle's free 5-day Agents intensive. A hands-on, guided course on building AI agents, with daily PDF guides; one of the better no-cost ways to get a team fluent fast.
  • Skip AI? You might be tripling your layoff risk. New Gallup data out in Bloomberg today: tech workers who use AI monthly had a 6% layoff risk. Those who rarely touch it, 18%. Three times higher.