Two developments that Dutch professionals are already feeling in practice: the watermarking obligation for generative AI is drawing closer, with the deadline on 2 December, and the three largest language models have now fully aligned their prices. Here are the facts at a glance.
Watermarking of AI output: deadline 2 December 2026
The AI Act has been in force for months, but pressure on providers of generative AI is concretely increasing in the coming weeks. Article 50 obliges providers to mark AI-generated text, images, audio and video in a machine-readable way so that detection is possible, the so-called watermarking obligation.
For systems that were already on the market before 2 August 2026, a transition period applies: the deadline shifts to 2 December 2026. That is less than eleven weeks away. Systems launched after 2 August had to have this built in at market introduction.
The technical methods recommended by the EU GPAI Code of Practice: visible overlays, invisible digital watermarks and metadata tagging via the C2PA standard. For organisations that publish AI-generated images or video, think marketing departments and media companies, this is not only a provider issue: internal publication processes must also be able to demonstrate what was created by AI.
Enforcement in case of breach: up to €15 million or 3% of worldwide annual turnover, whichever is higher. For more on what the transparency obligations mean for your use, see our safety page.
DeepSeek still does not publish a retention policy
In price comparisons, DeepSeek V4.1 Flash is becoming increasingly attractive for high-volume tasks: $0.30 per million input tokens and $1.20 output, with a half-price off-peak discount. That makes it roughly thirteen times cheaper than the frontier models.
But there is a substantive objection: DeepSeek does not publish a clear data deletion policy. Data is retained "for as long as necessary for the service and legitimate business interests", with no fixed period. Combined with storage in China and earlier blocks by Italian and other European regulators, this makes the model less suitable for processing involving personal data or confidential business information. See also our DeepSeek model profile for the current status.
Frontier models: convergence at $10/$50
GPT-6 Astra (launched 3 September 2026) and Claude Fable 5.1 (1 September 2026) are both priced identically: $10 per million input tokens and $50 per million output tokens. The difference lies in caching behaviour: Anthropic reduced the cache-read cost of Fable 5.1 to $0.25 per million tokens, which makes a significant difference for repeated long contexts.
Google Gemini 3.8 Flash is priced at $0.75 input / $3.75 output, but that promotional price only applies until 31 December 2026, after which the standard $1.50/$7.50 rate kicks in.
Choosing between these models depends largely on your workload. Analysis of multiple benchmark sources suggests a three-tier approach: high-volume tasks via DeepSeek or Gemini 3.5, mid-weight work via Gemini 3.8 Flash, and complex reasoning via Fable 5.1 or GPT-6 Astra. See the model comparison in our models overview.
What you can do now
- Check your tooling before 2 December: does your SaaS provider use generative AI for image editing or content creation? Ask them in writing how they are implementing the watermarking obligation.
- Evaluate your DeepSeek use: do you process personal data or contractually sensitive information? If so, the GDPR risk with DeepSeek is not hypothetical. Opt for a model with explicit EU data residency.
- Keep an eye on the promotional Gemini price: if you plan large-scale API use via Gemini 3.8 Flash, make sure your budgeting accounts for the doubling on 1 January 2027.
You can find the full EU AI Act deadline calendar and an explanation per obligation via compliancehub.wiki.