Qwen, DeepSeek, Anthropic: The Agent Battle Escalates

Today makes it pretty clear where the AI market is heading: away from “look what the model can do” and toward “look what it can complete autonomously over hours.” At the same time, security is getting more serious, the price war is intensifying, and AI is moving further into the work surfaces you already use every day. In short: more power, more risk, more everyday utility.

🤖 Qwen3.7-Max: Agent performance at a new level

Alibaba is releasing Qwen3.7-Max, a proprietary model that apparently isn’t just meant to provide good answers, but above all to reliably handle long, autonomous tasks. According to The Decoder, the model solved a complex development task in 1,158 autonomous steps — exactly the kind of patience test where many LLMs eventually start wandering off creatively. In benchmarks, Qwen3.7-Max is said to be at the level of Claude Opus 4.6 and to outperform Chinese competitors. The robotics angle is also interesting: the model was additionally shown as control software for a four-legged robot.

Why does this matter? Because a trend is becoming clearer here: the next major differentiation in the AI market is no longer just chat quality, but agent capability, tool use, and reliability across long workflows. For developers, that means autonomy is getting cheaper and more capable — but also harder to control. For the market, it means proprietary models from China are seriously entering demanding agent workloads.

🛡️ Anthropic warns: bugs are being found faster than they can be patched

As part of Project Glasswing with around 50 partners, Anthropic reports more than 10,000 critical vulnerabilities in mission-critical software. At first, that sounds like a security win — and it is. But it also reveals the problem: according to The Decoder, Claude Mythos Preview appears to find bugs faster than teams can fix them. Anthropic therefore explicitly warns of a transition period with elevated risk. Even more bluntly: no company has yet put in place safeguards against the misuse of such models.

This is an important reality check for the AI security debate. More automated vulnerability research does not automatically mean more safety in day-to-day operations. If offensive capabilities grow faster than patch processes, a dangerous window opens up. For companies, that means security teams need to speed up prioritization, triage, and patch management significantly. Or put differently: anyone who still thought “we’ll patch it later” just got a very expensive wake-up call.

💸 DeepSeek makes the price war permanent

DeepSeek is turning the screw on the market’s most uncomfortable lever: price. According to The Decoder, the previous 75% discount on V4-Pro is now permanent. At $0.435 per million input tokens, the model is said to be around 11 times cheaper than GPT-5.5 and even 29 times cheaper on output. This is especially relevant for token-intensive agent systems, i.e. exactly the kind of applications that require many steps, lots of context, and heavy model usage.

This is strategically explosive. If a capable model drops this sharply in price, Western providers come under pressure — not only on margin, but also on product strategy. After all, agents cost not just development effort, but ongoing tokens as well. If you’re too expensive here, many teams will simply find you unattractive. The effect is likely to resemble what happened in cloud computing: once a price anchor is set, other providers either have to follow or differentiate clearly through quality, tooling, or compliance.

📊 OpenAI Codex gets Appshots for more context

OpenAI is continuing to evolve its coding assistant Codex toward a “work assistant with context awareness.” With Appshots, Mac users can send the contents of any app window to Codex at the press of a button so the model gets more visual context for the task. The Decoder reports on this. It’s useful because many problems aren’t in the code itself, but in the interaction between the IDE, browser, documentation, or UI.

For you, that means AI tools are increasingly becoming context-sensitive copilots, not just text generators. That can save a huge amount of time, especially for debugging, reproducing errors, or tasks that otherwise involve jumping between multiple windows. At the same time, it naturally raises the question of what data is being sent where exactly. Convenience and control are rarely best friends — they tend to prefer a polite distance.

📽️ ChatGPT comes directly to PowerPoint

OpenAI is also moving further into the familiar tools used in everyday office work. A new ChatGPT plugin for PowerPoint can, according to The Decoder, create presentations from notes, documents, or images and edit existing slides. The beta add-in is available worldwide for all plans, but still does not support complex formatting. OpenAI also recommends backing up important decks in advance — advice that is rarely a great sign, but at least it’s honest.

The relevance is obvious: anyone who builds slides regularly gets a real productivity boost from AI. At the same time, the plugin shows that the next evolutionary step is not “yet another chat window,” but direct integration into existing work environments. That’s where the biggest leverage comes from: less copy-paste, less context switching, more immediate execution. Once formatting catches up too, this could become a real standard workflow for many teams.

⚡ Efficient tandem solar cells: industry gets closer

Not all good tech news is about LLMs. Researchers have, according to heise, developed a vacuum process that makes it possible to apply the perovskite layer for tandem solar cells evenly and at scale. That was previously a key bottleneck for industrial production. This brings the scaling of these especially efficient solar cells one step closer.

Why is this in the Daily Digest? Because energy and AI are becoming increasingly interconnected. More compute requires more electricity — and the more efficient energy technologies become, the easier it is to think of AI infrastructure and sustainable power supply together. So this is not a side note, but part of the larger infrastructure question behind the AI boom.

🧰 Tool tip of the day: test PowerPoint + AI properly

If you create a lot of presentations, the new ChatGPT plugin for PowerPoint is worth a look — especially if you want to turn notes into first-slide drafts quickly. My advice: use it for drafts, not for final board-level decks. For that, # is ideally combined with a good workflow for review, source checking, and layout control. If you often jump between documents, images, and slides, it can save you a lot of time — and hopefully also spare you the classic “where is the final_final_v7.pptx?” moment.

---

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