AI is becoming more expensive, more agentic, and more medical

Today, it’s becoming pretty clear where the AI market is heading: away from the friendly chatbot and toward expensive, specialized, and increasingly autonomous systems. At the same time, things are getting more serious in companies, in medical applications, and around privacy — exactly the areas where “let’s just try it out” is no longer enough.

For you, that means more productivity, more responsibility, and more pricing pressure. And yes, also more announcements that sound a bit like every vendor has just discovered the Holy Grail of intelligence.

💸 AI monetization is becoming a hard reality

The AI world is getting an uncomfortably important lesson in economics: models don’t just cost money to train, they mainly cost money to run. The Verge reports that providers like Anthropic and OpenAI now have to pay closer attention to how much compute their products consume — and how much of that cost they can later recover. This is especially true for agent workflows, which need multiple tool calls, longer contexts, and more compute time.

Why does this matter? Because the era of “we’ll scale everything first and figure out the bill later” is slowly coming to an end. For users, that could mean stricter limits, higher prices, or less generous free access. For companies, it’s a signal that AI integration must be calculated cleanly not just technically, but economically as well. The magic of the demo is still there, but the cloud bill still arrives.

🧑‍⚕️ OpenAI is building ChatGPT for doctors

OpenAI has introduced a free version for medical professionals with ChatGPT for Clinicians. It also comes with a new benchmark, which OpenAI says shows GPT-5.4 can outperform human doctors on clinical tasks — even when those doctors have unlimited time and internet access. That’s a pretty bold claim, even by AI standards.

The relevance is obvious: healthcare is one of the most sensitive and, at the same time, most promising AI markets. Here, it’s not just about efficiency, but about trust, liability, and real effects on patients. If systems like these prove themselves in practice, they could massively speed up documentation, differential diagnoses, and research. But benchmarks are not a clinic. And a model that shines on paper still has to prove in the real world that it doesn’t just sound smart, but is also safe.

🧠 GPT-5.5: OpenAI turns the agentic dial up

At the same time, OpenAI has announced another major model with GPT-5.5. The focus is clearly on agentic capabilities: the model is supposed to handle complex tasks independently across multiple tools. So not just answer, but plan, execute, and adjust — exactly what the whole industry is currently working toward.

The pricing is also interesting: according to the report, OpenAI is charging double the API price. That makes one thing very clear: “more intelligence” in practice often also means “more cost.” For developers, GPT-5.5 could be interesting for high-quality, business-critical workflows, but not necessarily for every standard use case. Agents are not just magic — they’re also tokens with ambitious price awareness.

🧩 Google is letting AI write 75 percent of new code

According to The Decoder, 75 percent of new code at Google already comes from AI — and is then reviewed by humans. That’s a remarkable figure, because it shows how far AI-supported development has already come in the everyday life of a Big Tech company. This is no longer about experiments in isolated teams, but about productivity in the core process.

What does that mean? Developer roles are increasingly shifting from “write everything yourself” to “suggest, review, integrate.” Humans become more like reviewers, architects, and quality controllers. That can massively increase speed, but it also brings risks: bad code doesn’t automatically become better just because AI wrote it, and debugging remains annoying. In short: AI makes software development faster — but not automatically easier.

🏓 Sony robot Ace plays table tennis at pro level

With Ace, Sony AI is showing a robot that, according to the company, can keep up at expert level in a sport for the first time. Table tennis is a tough benchmark for that: fast reactions, precise motor control, constant anticipation. Exactly the kind of task where robots traditionally start to sweat — figuratively speaking, of course.

The relevance goes beyond sport. Such systems are a testbed for robotics, perception, and real-time decision-making under uncertainty. What works in table tennis could later become valuable in industry, logistics, or assistive robotics. It’s a nice example of how research often starts with things that seem like a show at first glance — until it becomes clear that there’s some pretty serious technology underneath.

🔒 OpenAI releases open-source tool for redacting data

Privacy in everyday AI use is not optional, it’s mandatory. That’s exactly why the new open-source model Privacy Filter is interesting: it is designed to detect personal data in text and redact it. This can be especially helpful for internal workflows, logs, support tickets, or documents that end up in AI systems.

For companies, it’s a practical tool because it makes the balancing act between usefulness and privacy a little easier. Of course, a filter like this doesn’t replace a proper data strategy, but it is a useful safety net. Especially in regulated environments, the rule is: better redact before the prompt than explain afterward. That saves time, nerves, and usually a few very unpleasant meetings.

🛠️ Tool tip of the day: Privacy Filter for sensitive texts

If you regularly work with internal documents, customer data, or medical texts, it’s worth taking a look at the OpenAI Privacy Filter. The model identifies personal information and can automatically mask it before content is sent to other systems. For teams with privacy requirements, this is a useful building block in the AI workflow. #

Don’t want to miss any news? Subscribe to the newsletter