Two developments from 24–25 July that matter directly for Dutch AI professionals: Anthropic released Claude Opus 5 yesterday, and the final text of the Digital Omnibus makes clear what is really expected of you from 2 August 2026, and what is not.
Claude Opus 5: what you get
Anthropic launched Claude Opus 5 on 24 July as a "near-frontier model" that comes close to the flagship Fable 5, but at half the price.
Pricing and speed
The standard prices are $5 per million input tokens and $25 per million output tokens, the same as Opus 4.8. New is a fast mode that makes the model 2.5 times faster for twice the base price, aimed at time-critical agent tasks.
Effort levels (effort toggle)
Opus 5 offers three effort levels: low, medium and high. You choose whether to deploy the model quickly and cheaply for routine tasks, or give it more computing power for complex reasoning steps. That makes it usable across a wide range of workflows, from e-mail follow-up to multi-step legal or financial research.
Benchmarks
On CursorBench 3.2 (coding), Opus 5 sits within 0.5% of Fable 5, but at half the cost. On OSWorld 2.0 (computer use), the model outperforms all competitors except Fable 5, at a third of that price. Its ARC-AGI-3 score is three times higher than the best-scoring other model on that evaluation, Anthropic reports.
What this means for you
If you use Opus 4 or Fable 5 today for knowledge work or agentic tasks, Opus 5 is a sensible alternative to reduce costs without sacrificing much quality. Compare models on our model overview page.
GDPR/privacy caveat
For Opus 5, Anthropic does not specify any EU data residency or explicit GDPR processing instructions beyond the existing business terms. Business API customers do not have their data used for training; Free/Pro/Max users can opt out via account settings. Do not use personal data in the consumer service if you want to limit privacy risks. See also our Claude safety profile.
AI Act Omnibus: what is really on the agenda for 2 August?
Early reactions suggested that the Digital Omnibus postponed all AI Act deadlines. That is not entirely correct.
On 2 August 2026, the obligations for high-risk AI systems under Annex III do NOT lapse: these have indeed been moved via the Omnibus to 2 December 2027, a delay of 16 months, confirmed by the EU Council.
But Article 50 (transparency obligation) remains in force on 2 August. In concrete terms, this means:
- Every chatbot, voicebot or virtual assistant must identify itself as AI at the start of a conversation.
- AI-generated content (text, image, audio, video) must be labelled as synthetic where technically feasible.
- Systems applying emotion recognition or biometric categorisation must inform the person concerned.
Fines can reach €15 million or 3% of worldwide annual turnover. This applies to every organisation deploying AI systems towards end users in the EU, regardless of size. See also our overview of AI Act obligations by sector.
Sources
- Anthropic: Claude Opus 5 announcement
- Fortune on the Opus 5 effort toggle
- Gibson Dunn: EU AI Act Omnibus, postponed deadlines
- Cloud Security Alliance: AI Act Omnibus, 2 December 2027 deadline