· reviewed by René Mattis · vona
How the weekly review is made
Every Monday a vona tool reads the news from 18 selected sources. An AI (Claude by Anthropic) summarises them and carries the review over into English; an automatic check makes sure every link comes from these sources or from our own articles. Before publication, René Mattis reads the drafts, checks the key statements against the sources and approves them.
Coding Agents Tested, Meta Backtracks and Nvidia
Coding agents rarely solve difficult tasks completely, Meta withdraws new AI clauses, and Nvidia investigates its CEO's relationship.
Transparency: This review is produced with AI assistance. A tool collects the reports of selected sources, a language model summarizes them, and we check and contextualize them. The summaries are based on the publications of the linked sources. For details, it is always worth consulting the original.
A new benchmark shows how far coding agents still are from finishing real tasks. Meta has reversed a controversial change to its terms, and OpenAI has released a model for security testing. At Nvidia, the CEO himself is the subject of an investigation.
This week’s key stories
Coding agents deliver a lot, but rarely a finished result. Alibaba’s Co-Design Bench benchmark tests agents on real, difficult user requests. Even GPT-6 Astra, the top-ranked model, fully solves only 26.4 percent of the tasks. On older benchmarks, the same models score above 99 percent (The Decoder (opens in a new tab)).
Meta withdraws new terms of service for Facebook Pages. Page owners will not have to allow linked content to be used for AI training after all (Ars Technica (opens in a new tab)).
OpenAI introduces GPT-6 Cyber, a dedicated model for IT security. It is designed to find exploitable vulnerabilities and develop fixes, and it is trained to work only within an authorized scope (OpenAI (opens in a new tab)). OpenAI is also working with Hacktron on the security of its own systems, including testing connections to third-party providers (OpenAI (opens in a new tab)).
Nvidia’s board is investigating CEO Jensen Huang’s relationship with Alice Huang, an executive at Megaspeed. According to reports, she was recruited by China’s Ministry of State Security (The Verge (opens in a new tab)). In a memo to employees, Huang admitted to the affair (TechCrunch (opens in a new tab)). His future at the company remains uncertain.
Anthropic and OpenAI keep raising capital. According to the Financial Times, Anthropic still plans to go public before Thanksgiving despite the sell-off in tech stocks. It is aiming for a valuation of 1.25 to 1.75 trillion dollars (The Decoder (opens in a new tab)). OpenAI has reportedly raised another 11 billion dollars in an ongoing funding round, at a valuation of 1.25 trillion dollars (TechCrunch (opens in a new tab)).
What this means for businesses
Agents need review and clear processes. The benchmark confirms what many people see in their daily work: code written by agents is a draft, not a finished product. Simon Willison uses pull requests as his unit of work to group and review changes from Claude Code or Codex (Simon Willison (opens in a new tab)). For ways to put this into practice in your team, see AI agents in everyday work and our Git workflow.
More options for keeping control of your own data. t3n tested Anything LLM, a tool for running AI locally with web search and knowledge from your own documents (t3n (opens in a new tab)). Mistral is promoting Studio with the promise that you keep full control over models, data and infrastructure (Mistral AI (opens in a new tab)). Before you decide on a solution, read AI and data privacy.
Your content and AI training. Meta’s reversal shows that platforms can change their rules on how content is used, and that they do. If you keep track of where your content appears and what it may be used for, you can make informed decisions. Related reading: GEO: How to become visible in AI answers and our review covering the opt-out for AI crawlers.
On the rise
- Grok: According to Google Trends, search interest in Germany rose by about 30 percent compared with the previous week.
- Anthropic: According to Google Trends, searches combining it with “künstliche Intelligenz” (artificial intelligence) were about 180 percent more frequent than in the previous week.
More from the sources
- Anthropic explains how Claude decides what to remember and how “dreaming” is meant to help (Anthropic (opens in a new tab)).
- t3n briefly tested live translation in ChatGPT, DeepL and Google. All three have their quirks (t3n (opens in a new tab)).
- Over one weekend, Andrej Karpathy used Claude Code to build HackerTube, a curated video platform based on Hacker News comments (The Decoder (opens in a new tab)).
If you want to know what any of these developments means for your project, get in touch with us at vona.