Google has a new top model, but you cannot use it yet. And if you use AI in recruitment and selection, you have had obligations since August, even though the most demanding rules have been postponed. This is what you need to know.
Gemini 4 Argon: announced, not yet available
On 30 September, Google announced Gemini 4 Argon, a model for software development, knowledge work and cyber security. For now, only a select group of cyber defenders will get access through Google's Fairwind programme. After that, paying API customers and Google AI Ultra subscribers will follow "as soon as possible". Google does not give a date.
The price is in two stages. During the introductory period you pay $2 per million input tokens and $10 per million output tokens. After that it becomes $4 and $20. That second rate matches what Claude Opus 5.5 costs; the introductory price matches GPT-6.1 Sol and Claude Sonnet 5.5. New is the output limit of 1 million tokens, compared with 64,000 for the previous model. This means the model can write very long documents or large amounts of code in one go.
Google cites strong benchmark scores, such as 77.9% on DeepSWE for software work. These are the provider's own figures. According to Bloomberg, via Investing.com, some Google staff doubt whether the model is genuinely good at practical programming work. Google disputes this. Argon does not yet appear on the independent LMArena leaderboard. The best Gemini model there is currently Gemini 3.8 Flash, in position 10 with 1492 points; Claude Opus 5.5 leads with 1509.
For you, this means nothing yet. If you work as a software developer and want to test Argon once it is available, do so alongside your current model on your own tasks, not on the basis of benchmarks. You can read how Google handles your data at /gemini/.
EU AI Act and recruitment: what already applies
Recruitmenttech.nl set out what the EU AI Act currently requires of recruiters. The essentials:
- Since 2 August 2026, the transparency rules apply. If a candidate interacts with a chatbot or another AI system, they must be told so. In online assessments involving AI, it must also be clear that AI is being used.
- Since February 2025, you must ensure that staff working with AI have sufficient understanding of it (AI literacy).
- The strict rules for high-risk AI in recruitment only apply from 2 December 2027. This was postponed through the European AI omnibus, which entered into force on 27 July 2026. It concerns systems that filter, score or rank candidates, such as a tool that labels CVs as a "good match" or "poor match".
The postponement gives you time, but not an exemption. Map out now which tools in your recruitment process assess candidates. That way you will know which vendors to ask for documentation later. In the meantime, GDPR continues to apply in full to all candidate data.
Do you work in recruitment, HR or the staffing sector? On those pages we compare the tools you use.
What you can do now
- Hold off on making plans around Gemini 4 Argon until it is actually available, and budget for the price after the introductory period.
- Tell candidates when they are dealing with a chatbot or AI assessment.
- Make a list of recruitment tools that score or rank CVs.
- Compare models on price and data storage via /modellen/ and /veiligheid/.