Two developments of the past few weeks directly affect Dutch organisations: the EU has postponed the high-risk obligations of the EU AI Act, and one of the most capable open-weight models available right now, GLM-5.3 from China's Z.AI, has released a quiet but notable update. What does this mean for your work?
AI Act: what has been postponed, and what has not
On 27 July 2026, the so-called Digital Omnibus took effect (EU 2026/1744). It shifts the deadline for high-risk AI systems considerably:
- Standalone high-risk systems (Annex III): from 2 August 2026 to 2 December 2027
- High-risk systems in regulated products (Annex I, such as medical devices and machinery): from August 2027 to August 2028
- AI sandbox obligations (Article 57): from August 2026 to August 2027
That sounds like breathing room, but there is an important exception: the transparency obligations under Article 50 still apply from 2 August 2026. National market surveillance is already active. In practice, this means that if you deploy a chatbot or other AI system that communicates directly with people, it must be clear that it is AI, before the first interaction. Synthetically generated content must be labelled.
What already applied and continues to apply: the ban on prohibited AI practices (Article 5, in force since February 2025) and the AI literacy duty for staff (Article 4).
The practical question for many organisations: which systems already fall under the law? That depends on the role you play (deployer, provider, user) and the application. The checklist published by the CFO editorial team (inventory your systems, establish roles, document governance, monitor outcomes) is a useful starting point.
See also our safety page for an overview of how the major models score on GDPR compliance.
GLM-5.3: quiet upgrade, noticeable leap
On 14 August 2026, Z.AI (the Chinese Zhipu AI) released GLM-5.3. Technically, it shares the same 743 billion parameter base as the previously released GLM-5.2, but it stands out through more intensive post-training. The result: on Terminal-Bench 3.0 the score rises from 4.6 to 28.3, an improvement showing that the model has become considerably better at carrying out complex tasks in a terminal environment.
The price remains the same as GLM-5.2: $1.40 per million input tokens and $4.40 per million output tokens. For context, that is comparable to DeepSeek and considerably cheaper than GPT-5.x or Claude Sonnet 5.
GLM-5.2 scores 62.1 on SWE-bench Pro, the highest score among open-weight models, ahead of GPT-5.5. GLM-5.3 builds on the same foundation. For development teams that want to use open-weight models for software development or terminal-based tasks, this is an option worth serious attention.
One point of attention for Dutch organisations: Z.AI is a Chinese company and data processing does not take place within the EU. That makes it less suitable for processing personal data or sensitive business information. See our model overview for a comparison of hosting and privacy profiles.
What to do this week
- Check whether your chatbots and AI tools meet the transparency requirements (Article 50): users must know that they are communicating with AI.
- Make an inventory of your AI systems: do they fall under Annex III or Annex I? If so, you have until December 2027 or August 2028 to achieve full compliance.
- When using open-weight models such as GLM-5.3: make sure the processing of personal data always takes place within the EU, or use an EU-hosted deployment.
For a sector-specific overview of which tools and models have been assessed for GDPR and the AI Act, see the sector pages on betin-xt.com.
Sources: CFO.nl – AI Act August 2026 · Niels Roest – postponed deadlines · Kingy.ai – GLM-5.3 specs · MorphLLM – GLM-5.2 benchmarks