OpenAI, Anthropic and Google turn up the AI heat
OpenAI, Anthropic and Google deliver new AI features for coding, images and agents. Plus: a possible math breakthrough, TSMC records and a Windows zero-day.
Inhaltsverzeichnis
Today it’s once again all about the usual AI chaos: OpenAI sharpens Codex, Anthropic follows up with Claude Opus 4.7, and Google makes Gemini more personal than some people would like. On top of that, there’s a possible math coup by GPT-5.4 Pro, fresh records at TSMC, and unfortunately also a new Windows zero-day.
In short: product velocity, a race around developer workflows, more personalized AI — and a security hole you should not ignore.
🤖 OpenAI makes Codex significantly more powerful
OpenAI is giving its developer tool Codex a major upgrade, positioning it even more clearly as a response to Anthropic’s Claude Code. According to The Decoder, Codex can now not only help with programming better, but also operate a Mac independently, generate images, and remember preferences. Especially interesting: the tool is said to be able to keep working on tasks for weeks instead of clocking out after a few prompts. That’s a real step toward agentic coding, meaning AI as an active work assistant rather than just autocomplete with delusions of grandeur.
Why does this matter? Because right now the industry is deciding what software development will look like in the next few years: who can delegate tasks robustly, who can understand workflows, and whose tool becomes the default inside the developer IDE? OpenAI is directly challenging Claude Code’s momentum advantage. For you, that means the AI development environment is becoming much more autonomous — and therefore more useful, but also more in need of control. Because once an AI is allowed to operate your Mac, trust is no longer a footnote.
🧠 GPT-5.4 Pro is said to have solved an open math problem
A genuinely remarkable research claim also comes from OpenAI: According to The Decoder, GPT-5.4 Pro is said to have independently solved an open Erdős problem in around 80 minutes. Terence Tao, one of the most renowned mathematicians of our time, apparently sees this as a meaningful contribution to mathematics. This is not “AI did a homework assignment,” but rather: a model worked on a problem that is non-trivial in the academic world.
Why is that important? Mathematics is a hard test bed for AI, because it depends on logical reasoning, clean proofs, and precise chains of inference. If a model really produces new results here, that would be more than a benchmark gimmick. It would show that LLMs can not only reconstruct text, but also support real research in limited contexts. At the same time, as always: a single success is not proof of general superintelligence. But it is a pretty good reason to look more closely. And yes, the idea of a language model untangling an old math knot is deliciously ironic in a very dry way.
🔐 New Windows zero-day grants admin rights
On the security front, the news is less good: According to heise, there is a new unpatched Windows zero-day that, rather inconveniently, leads to system privileges via unsafe behavior in Windows Defender and a file API. That is especially nasty because the attack can escalate local privileges — from a normal user account to administrator rights.
Why you should take this seriously: in AI-supported workflows, tools often run with elevated privileges, access many files, or are deeply integrated into development environments. A Windows zero-day is therefore not just a classic IT security problem, but also a risk for modern AI workstations. Once an attacker has admin rights, the path to data exfiltration, manipulation, or lateral movement across the network is not far. Bottom line: watch for updates, harden systems, minimize privileges. Security romanticism is not a strategy.
📈 TSMC delivers records thanks to AI demand
Chip giant TSMC is reporting new record figures, according to heise — and the main driver remains the extremely strong demand for AI chips. Particularly notable: even geopolitical uncertainty, including the Iran crisis, apparently isn’t causing any noticeable dent. TSMC speaks of “extremely robust” demand while shipping more wafers and enforcing higher prices.
For the AI market, that is an important signal. It shows that the boom is not just a software phenomenon, but continues to put pressure on the physical supply chain. If you want more AI, you don’t just need models and benchmarks — you also need expensive manufacturing, packaging, and reliable capacity. As long as TSMC’s order books stay this strong, one thing is clear: the hunger for compute is far from satisfied. Or put differently: the AI revolution runs on chips, and those chips currently cost more than nerves.
🖼️ Gemini pulls images from Google Photos
Google is making Gemini even more personal: According to The Verge, the AI can now use data from Google Photos and other apps to generate personalized images. With the “Personal Intelligence” feature, prompts like “Design my dream house” or “Create a picture of my desert island essentials” are no longer supposed to produce generic AI imagery, but images based on your real context. Google is also combining this with its image model “Nano Banana 2”.
This is exciting because it pushes the boundary between assistance and context exploitation even further. If an AI knows what you photographed, saved, or researched, it can deliver much more relevant results. At the same time, the question of privacy and consent becomes even more important. Personalization is convenient — but it only works if the AI knows quite a lot about you. For users, that can be useful. For privacy advocates, it’s more like a slight twitch of the eyelid.
🧪 Anthropic releases Claude Opus 4.7
Anthropic is countering as well, launching Claude Opus 4.7, its strongest generally available model to date. According to the report, the new Opus is a step forward for advanced software engineering tasks, especially complex coding scenarios that previously needed more handholding. It is also said to be better at analyzing images, following instructions, and being more creative when creating slides.
The takeaway is clear: the competition between OpenAI and Anthropic is no longer just a model race, but a fight for the best work assistant for knowledge work and software development. Whoever leads in coding, multimodality, and instruction quality gets the better use cases, the more loyal developers, and ultimately probably also the paying teams. For you, that means the next wave of AI tools will do less “chatting” and more “working.” And that is exactly where it will be decided who really helps in everyday life — and who just sounds impressive.
🛠️ Tool tip of the day
If you’re experimenting with agentic coding yourself, today is a good day to look at modern dev assistants that go beyond simple code completion. Especially interesting are tools that can read files, plan tasks, and maintain longer work streams. That helps you better understand where Codex, Claude Code, and similar systems are heading.
Recommendation: #
Don’t want to miss any news? Subscribe to the newsletter