Microsoft Build 2026 declared the end of the app era, Anthropic made honesty the selling point of its newest model, and behind the scenes the same question kept popping up everywhere: what does all that AI actually cost? An edition about grand promises, incoming bills, and a backlash that's getting louder. Every two weeks on the Xylos blog, we bring you a sharp and honest overview of what's really moving in the world of generative AI, with the context you need.

The foundation remains the biweekly LinkedIn overview by Tom Van 't veld, Learning Innovator at OASE (powered by Xylos). Tom follows AI developments closely, and we translate his observations into what they concretely mean for organizations and the people working in them. Welcome to edition five.

On June 18, AI celebrates its seventieth birthday. On that date in 1956, the Dartmouth Summer Research Project on Artificial Intelligence kicked off, the moment the term artificial intelligence was formally coined for the first time. What once started as a modest brainstorm about chess programs is now writing its own codebase. Below are five stories from the past two weeks that matter most for your organization.

Microsoft Build 2026 centered on a key message from Satya Nadella: the era of apps and operating systems is over. Scout, the new always-on Copilot agent, autonomously monitors inbox, calendar and Teams sessions, even without the user asking a question. Project Solara sketches the next step, with a platform that runs AI agents instead of traditional applications, built on Android instead of Windows.

Practice calls for realism. Microsoft 365 Copilot got a speed boost and a refreshed design. ZDNET tested the paid Copilot agents and found they came across confidently, but hadn't yet reached the expected results for more complex tasks. Agent technology is clearly still very much maturing.

The vision is grand, reliability remains the pain point for now. For organizations, agents are meanwhile shifting from pilot to daily reality, and that reality is messier than the keynote suggests. Pick use cases where an agent demonstrably saves time, and build in governance before the tooling grows faster than your control over it.

Anthropic launched Claude Opus 4.8, with honesty as a notable selling point. The model is roughly four times less likely to silently let errors in generated code slide, and more often flags when it doesn't know something with certainty. Full trust still isn't warranted, though, since the model can hallucinate with a lot of confidence, and you only notice once you check the output.

Safety was more broadly on the agenda too. OpenAI introduced Lockdown Mode for ChatGPT, an optional setting that disables web search, Deep Research and Agent Mode to protect against prompt injection, where malicious instructions are hidden inside web pages or uploaded files. Anthropic added to that with a free security plug-in for the Claude Code terminal that detects vulnerabilities in codebases.

For anyone working with sensitive information, the choice of which model to use for which use case becomes decisive. Models that dare to show their doubt lower the risk, though they don't replace human review. Put clear agreements in place around data and review, so the gains in speed don't get eaten up by fixing errors.

Update after Tom's overview: on June 9, Anthropic made [Claude Fable 5](https://www.anthropic.com/news/claude-fable-5-mythos-5) generally available, its most capable public model to date, available among others through the Claude API and Microsoft Foundry. For sensitive domains such as cybersecurity, the model conservatively falls back to Claude Opus 4.8.

The financial side of AI hit hard. On June 1, GitHub switched from a fixed monthly fee to token-based AI Credits. The base price per seat stays the same, but anyone using agents intensively sees costs climb sharply. The irony is striking: the same companies that once encouraged their employees to use as much AI as possible are now presenting the bill based on exactly that usage.

That the numbers can spiral out of control was shown by a few examples. One anonymous company accidentally spent 500 million dollars on Claude in a single month, after nobody had set a usage limit on employee licenses. Microsoft pulled back its own internal Claude Code licenses, and Uber burned through its entire 2026 AI budget by April already. A Fortune article citing internal Microsoft research even claims that, in a range of scenarios, AI tools cost more than human employees for the same tasks. Gartner warns that falling token prices won't offset rising consumption.

AI costs deserve the same discipline as any other IT expense. In a Microsoft 365 environment with a Premium Copilot license, this plays out less dramatically for now, but anyone setting up pay-per-use agents for colleagues without a license quickly runs into the same surprises. Set usage limits, track consumption, and tie every use case to an expected return before scaling up.

The reaction against AI overkill is gaining ground. Firefox's full redesign includes a button that turns off all AI features in one click, and DuckDuckGo is making its AI-free search more prominent as its usage grows. Criticism of Google's AI search mode keeps piling up: the search term is sometimes literally ignored, and the answers increasingly sound like advertising in tone. A study of 846,000 searches confirms that AI Overviews significantly reduce click-throughs to source pages.

On the other hand, attention to detection and transparency is growing. YouTube now automatically labels AI-generated videos, Google's SynthID watermarking technology is expanding to more formats, and Microsoft Clarity shows website owners which AI queries lead to their content.

For B2B organizations, search behavior is shifting, and with it, how customers find you. AI models summarize what they read on your website, so valuable visitor traffic is tipping. Invest in clear, authoritative content that still holds up inside a summary, and start making agreements today about labeling your own AI output.

The AI labs showed their most contradictory side. Anthropic published a call for a global pause in frontier AI development, backed by its own figures: Claude now writes eighty percent of its own codebase. Four days later, the company confidentially filed its IPO paperwork. In the United States, Florida became the first state to sue OpenAI, with Sam Altman personally named as co-defendant, while that same Altman lobbies in Washington against prior government approval for new models.

Meanwhile, Europe is looking for its position. Germany's SPRIND launched the Next Frontier AI competition, worth 125 million euros to build European frontier labs. An AWS study immediately sketches the Belgian paradox: with 62 percent AI adoption, our companies score above the European average of 54 percent, while 36 percent of startups are considering leaving Europe due to a lack of venture capital. Good at starting up, less good at holding on.

Beneath all that strategic noise, a quieter insight is emerging. Universities acknowledge that AI use has become the norm, while research shows that students who rely on AI as their primary teacher learn more superficially and develop less critical thinking ability. That same tension plays out on every work floor where AI arrives: the technology scales faster than the judgment needed to evaluate its output.

The rules of the game around AI are being written over the coming months, and your safest investment lies with your own people. AI literacy is becoming a baseline skill at every level. That's exactly where it overlaps with what Tom and his colleagues at OASE and Xylos Learning build every day: literacy that scales alongside the technology you bring in.

We'll be back in two weeks with the next edition.

About Tom Van 't veld

Tom has worked for Xylos for years, where he started as a Microsoft Office trainer and grew into the driving force behind innovative learning concepts. He is at the origin of OASE, Xylos's online learning platform, and developed, among other things, the Digital Coach concept, a Microsoft Teams Escape Room app, and the mAindset game that helps employees learn to prompt with AI in a playful way. As Learning Innovator, his focus in recent years has increasingly turned to what AI means for the way we learn and work. Want to react or keep the conversation going? Find him on LinkedIn.