Three developments from the past two days that you can use as a Dutch professional: a registration deadline for the Dutch Data Protection Authority pilot that expires on Monday, a substantially cheaper multimodal model from China, and a legal configuration of GPT-6 that delivers less for Dutch lawyers than the headlines suggest.
Registration for the FRIA pilot is open until 21 September
The Dutch Data Protection Authority is starting a pilot in October with roughly ten organisations on the FRIA, the fundamental rights impact assessment prescribed by the AI Regulation. Registration is possible until Monday 21 September; the pilot runs until mid-December, with a closing meeting in February.
The difference from a DPIA matters. A DPIA looks at privacy risks when processing personal data; a FRIA looks more broadly at the consequences of a high-risk AI system for all fundamental rights, including equal treatment and non-discrimination alongside privacy. Public authorities and parties delivering public services will have to carry out that assessment mandatorily from 2 December 2027.
The DPA stresses that the pilot is explicitly not an audit, inspection or enforcement process. It is meant to learn what organisations need. If you work at a municipality, implementing body or housing association and use algorithms in decision-making, this is an inexpensive way to find out in good time what is expected of you. See also our page on AI in government.
Qwen3.8-Omni-Flash: everything in one model, but only via the API
Alibaba's Qwen team released Qwen3.8-Omni-Flash on 18 September, with a context window of 1 million tokens. The model accepts text, image, audio and video as input and returns text. The pricing is striking: $0.15 per million input tokens and $0.47 output, with cache hits at $0.016. According to Alibaba, audio input is more than 98% cheaper per hour than with its predecessor.
Two caveats before you plan this in for transcribing meetings or customer calls. First, no open weights have been announced. The ability to self-host Qwen models is normally the GDPR argument in favour of this family; that argument falls away here. Second, the international Qwen services run through Alibaba Cloud in Singapore, with no EU data centre. For audio containing personal data (call recordings, patient contact, job interviews), that means a transfer outside the EEA that you must be able to justify. What that means in practice is set out on our safety page; the full Qwen profile is on the models page.
Astra for Law: US law, not Dutch
OpenAI launched Astra for Law on 17 September. Important to know: this is not a new model, but GPT-6 Astra with a legal search index and instructions layered on top. That index covers US case law, legislation, regulations and procedural rules across more than 230 million URLs, including case law from CourtListener by the Free Law Project.
On Vals AI's Legal Research Bench it achieved 54% correct answers across 200 questions. That is measurably better than a bare model, but it is still wrong in nearly half of cases, and the underlying data is American. For Dutch legal sources, nothing changes. For tools that do work with Dutch case law, see AI for lawyers.