AI Weekly Review CW 26: Export Ban, Acquisitions, GLM-5.2, EU AI Act Drift

Three topics from the second-to-last week of June deserve attention, because they show how quickly assumptions from 2025 are dissolving in 2026: the US export ban on Anthropic Fable 5 and Mythos 5 with the subsequent partial reversal, SpaceX acquiring the AI coding tool Cursor for USD 60 billion, and the release of GLM-5.2 as the new top model in the open-weight segment. Add the ongoing discussion about the EU AI Act high-risk deadline of 2 August 2026, which the Digital Omnibus package could postpone by up to 16 months.

1. Anthropic Fable 5: global launch, then ban, then partial restoration

On 9 June 2026, Anthropic shipped Claude Fable 5 and Claude Mythos 5, a new tier above Opus 4.8. On 12 June, the US Department of Commerce issued an export control directive citing national security concerns. Anthropic could not distinguish US users from foreign nationals in real time and disabled both models globally. On 18 June, global API access to Claude Fable 5 was restored with additional safety filters and identity checks; Claude Mythos 5 remains restricted to vetted defence organisations under Project Glasswing.

What this means for mid-sized decision-makers:

• Teams running business-critical processes on cloud AI models should expect temporary, unannounced model shutdowns from now on and run a second model option in parallel.
• US cloud AI is no longer a stable foundation when geopolitical tensions drive availability risk.
• Vendors whose business model depends on US cloud models need an exit strategy, including self-hosting or European alternatives.

2. SpaceX acquires Cursor for USD 60 billion, OpenAI buys Ona

On 16 June 2026, SpaceX acquired Anysphere, the parent company of the AI coding tool Cursor, for USD 60 billion. Through xAI, SpaceX now controls its own coding tool, its own model family, and its own distribution channel into the developer community. In parallel, OpenAI acquired Ona to strengthen the Codex platform.

The strategic read is clear: coding tools are becoming strategic platform assets, not applications. Teams that licensed Cursor or Codex as a tool over the past two years now have a relationship with a vendor that is part of a hardware, rocket, or model conglomerate. Lock-in risks shift, support and pricing strategies can change faster than expected.

For tool selection in mid-sized companies: keep a second tool ready, keep contract terms short, test self-migration paths.

3. GLM-5.2 from Z.ai: open weight closes the gap to the frontier

On 16 June, Z.ai released GLM-5.2 under an MIT licence, a 753B parameter Mixture-of-Experts model with a one-million-token context window. On the Artificial Analysis Intelligence Index, GLM-5.2 leads with a score of 51, ahead of the previous open-weight leaders DeepSeek V4 Pro and Kimi K2.6.

The shift is not just a new top score, it is a signal: open-weight models have closed the gap to proprietary frontier models far enough that they are the economically sensible choice for a broad class of use cases. For mid-sized software teams that stayed with OpenAI or Anthropic for compliance reasons, the MIT licence of GLM-5.2 shifts the argument. Self-hosting on own hardware becomes realistic once distillation pipelines are in place.

Three points to frame this:

• MIT licence means full commercial use without constraints, unlike many other open-weight licences with reach restrictions.
• 753B parameter MoE models are non-trivial to run: they need several GPUs, specialised inference software, and enough RAM. Self-hosting is a project, not a weekend cheat sheet.
• Benchmarks are indicators, not a substitute for own evaluation. Teams moving GLM-5.2 into production should test their specific use case against the model.

4. EU AI Act: the 2 August 2026 high-risk deadline is wobbling

The EU Commission tabled the Digital Omnibus package in May 2026, which could postpone the full application of the EU AI Act high-risk rules by up to 16 months. The original deadline of 2 August 2026 would mathematically shift to 2 December 2027. Formal adoption by Council and Parliament is still pending.

As long as the procedure runs, the legal situation stays unsettled. Mid-sized companies running AI-supported processes in the categories of HR, creditworthiness, critical infrastructure, or biometric systems should, in parallel to the ongoing discussion, have completed the inventory of which of their own use cases could fall under Annex III. Teams that do the inventory now are not in a rush in September.

What to watch in the coming week

The Anthropic shift shows that geopolitical risks for cloud AI are real in 2026. The Cursor acquisition shows that coding tools are turning into platform assets. The GLM-5.2 release shows that open-weight models are competitive for more and more use cases. The EU AI Act drift shows that regulation is not set in stone.

If you have one hour next week: check which of your cloud AI processes have a second model option. Anyone who runs two models side by side in 60 minutes is significantly better prepared for the next export ban than anyone betting on one card.