LoginMedia & News
The most important AI news from Switzerland and around the world, filterable by date range and topic. Members have access to the condensed news database in the members-only area.
Filter by date range and topic. The news articles link to their original sources. You can find swissAI's own posts on the blog.
Austria will introduce AI as a school subject starting in 2027, but the real problem remains: Schools must not only teach AI, but also fundamentally rethink learning.
To the SourceThe main advantages of the new GPT-Live-1 language model lie primarily in more natural and intelligent conversation flow, as well as new real-time features: • Natural conversations without disruptive interruptions • Greater intelligence through smart task delegation • Simultaneous live translations: One of the most remarkable new features is the ability to translate in real time • Wide availability: OpenAI is making the update available worldwide
To the SourceVertus AI claims to have developed the world's first nature-inspired artificial intelligence with a dynamic, brain-like architecture. Unlike conventional language models, which use static neural pathways, its system forms new neural connections for each query.
To the SourceAnthropic is expanding its AI agent, Claude Cowork, from the desktop app to the web and mobile. Users can start, monitor, and share tasks across devices while Claude continues to work in the background.
To the SourceSalesforce plans to invest $1 billion in Switzerland's AI infrastructure. This underscores Switzerland's growing importance as a European hub for AI.
To the SourceReports of data breaches in Switzerland have more than doubled compared to last year. Chief Data Protection Officer Adrian Lobsiger sees this as a sign of the public’s growing need for self-determination and control over their own data.
To the SourceYou’re comparing the price per million tokens. That’s exactly why you’re paying too much. Everyone is arguing about which model is the smartest. But the question that really matters gets lost in the mix: How many tokens does a model need to get there? That’s exactly where the wheat is separated from the chaff. The biggest cost driver is reasoning tokens—that is, what a model consumes to think for itself. Each round of thinking feeds back into the prompt for the next step. → A model that thinks extensively pays this price multiple times. The numbers are staggering. OpenAI’s GPT-5.5 Medium solves a Deep-SWE task with around 20,000 tokens. The best Gemini model needs around 270,000—for about half the score. With Anthropic’s Sonnet 5, the difference is even more pronounced. Where GPT-5.5 Medium managed with about 5,000 tokens, Sonnet 5 consumed a whopping 69,000. In Artificial Analysis’s large-scale test, it became the most expensive model ever tested at around $6,000, even ahead of Fable 5 at around $5,600. → Despite a lower starting price, it ended up being the most expensive option. The point is simple. If you only look at the price per million tokens, you’re comparing the wrong numbers. What matters is the cost per task solved.
To the SourceOpenClaw, a platform that has primarily been desktop-oriented until now, has launched its own app for smartphones. This gives users mobile access to its services.
To the SourceWith DSpark, Deepseek has introduced a new method that increases response speed per user by 60 to 85 percent. This is made possible by a combination of speculative decoding, batch token verification, and dynamic adjustment of computing capacity.
To the SourceThe Zurich Commercial Court has largely dismissed Palantir’s lawsuit against the online magazine “Republik.” Palantir has now accepted the ruling and will not appeal it; “Republik” is also unlikely to pursue the case further. “Republik,” in collaboration with the investigative collective “WAV,” had reported on Palantir’s business model, which relies on AI-powered data analysis and whose algorithms are considered trade secrets. Reporters Without Borders views the decision as an important signal for international press freedom. The ruling is expected to become final after the appeal period expires in mid-July 2026.
To the SourceOn June 30, 2026, Anthropic launched Claude Sonnet 5—the new standard model for Free and Pro plans. It is significantly more agent-like than Sonnet 4.6 and comes close to Opus 4.8 in terms of planning, tool usage, and multi-step tasks, but at a significantly lower cost (introductory price of $2.10 per million tokens through August 31, then $3.15). According to Anthropic, Sonnet 5 was not specifically trained for cybersecurity tasks and exhibits less undesirable behavior overall than its predecessor. For companies using Claude productively, this means more agent-like capabilities at the same price as previous versions of Sonnet.
To the SourceThe GitHub repository ai-swiss/base serves as a publicly accessible foundation for AI initiatives in Switzerland. It is available to the community as an open-source project.
To the SourceNewsletter
Every Friday, a concise summary of curated AI news from the swissAI community. Free and you can unsubscribe at any time.
AI Newsletter
Association members curate "AI Week." We summarize the highlights and send you the most important news, tools, and studies every Friday. It's free, and you can unsubscribe at any time.
