AI Blog
· daily-digest · 6 min read

Anthropic, Google and the New AI Head-to-Head

Anthropic, Google, and data brokers are setting today’s AI agenda: from agent swarms to robotics, training data, and chip porting.

Inhaltsverzeichnis

Today is about more than just the next model update: We’re seeing how AI is penetrating deeper into real work, infrastructure, and supply chains. At the same time, it’s becoming clear that the competition is not only about benchmarks, but about chips, data, costs, and control.

In short: Anyone who only looks at model performance today is missing the real story. It’s about who makes AI production-ready — and who keeps control over hardware, data, and processes in the process.

🤖 Anthropic Fable 5.2 cracks real freelance gigs

According to the Remote Labor Index, Anthropic’s Fable 5.2 completes 41.3 percent of real freelance jobs at professional quality — more than twice as much as the next-best systems, Opus 5.5 and GPT 6 Astra. This is not a minor benchmark footnote, but a pretty clear sign that AI is moving closer to “actually useful” than “impressive in the demo video” for certain knowledge and creative tasks. Especially interesting: a year ago, the top result was only 2.5 percent. So the progress is massive. At the same time, weaknesses remain visible in areas like product design and animation. What this means for you: AI can save more and more real working time, but it still cannot take over every creative workflow equally well. For companies, the question is becoming less whether AI can do tasks, and more which tasks can already be sensibly automated. Source: The Decoder

💼 Anthropic leadership cashes in big on the IPO

According to TechCrunch, each of Anthropic’s seven co-founders plans to make about $250 million in the IPO through an equal stock sale. That sounds like a typical “startup grows up” story, but it also sends a signal: Anthropic is no longer just a research and product marvel, but a company with massive capital interests and a clear market position. The fact that all seven founders are selling the same amount suggests a carefully balanced power and ownership structure — and that the team still wants to retain control. For the market, this matters because Anthropic is now firmly entering the league of major AI platforms that not only build models, but also serve capital-market logic. For users and enterprise customers, that usually means more stability, but also more pressure on monetization, pricing, and margins. Source: TechCrunch

💳 Antitrust authority scrutinizes PayPal terms

Germany’s Federal Cartel Office is taking a closer look at PayPal’s terms of use because they could be problematic from a competition standpoint. Specifically, the question is whether the payment service’s requirements distort competition and bind merchants too tightly to PayPal. This is a classic case of “convenience vs. market power”: PayPal is convenient for users and shops, but that convenience can eventually turn into dependence. For the AI context, this is indirectly relevant because many AI tools, SaaS products, and digital platforms run on the same payment and checkout infrastructure. If the rules there change, that quickly affects pricing, reach, and international expansion. Also noteworthy is the political undertone: European authorities are increasingly looking more closely when digital gatekeepers set their terms too aggressively. Source: heise

🧠 Google launches Gemini 4 Carbon with agent swarm

Google is making Gemini 4 Carbon generally available after a testing phase and is packaging the model with a two-part promise: better benchmark performance and lower costs. Especially notable is the new /swarm function, which lets multiple agents work together in a coordinated way on a complex task. That’s exciting because it further strengthens the trend from a single chatbot toward real AI teams. Instead of throwing one prompt into a model, you may soon distribute tasks among specialized agents that research, plan, and execute in parallel. For enterprise workflows, that could be a major lever — provided orchestration, error control, and security keep up. Google is also sending a price signal: good performance alone is no longer enough; it has to be efficient too. Welcome to the era of the agent economy. Source: The Decoder

🦾 DeepMind brings robotics to the Mercedes line

Google DeepMind is testing Gemini Robotics 2.5 with Apptronik for the first time in a real production environment at Mercedes-Benz in Hungary and Germany. The humanoid robot Apollo 2 sorted metal parts there with a reportedly 96 percent success rate. That’s remarkable because robotics often fails against harsh reality: lighting conditions, variability, grasp points, speed, exceptions — the realities of production are rarely as clean as the lab bench. If an AI-powered control model actually works reliably here, that’s an important step for industrial automation. For manufacturing, this doesn’t immediately mean “robots replace humans,” but it does mean less monotonous work, better flexibility, and new requirements for process design. Another sign that AI is not just writing text, but slowly touching the physical world. Source: The Decoder

⚙️ Anthropic ports Claude to AMD GPUs in three days

Anthropic claims that, with the help of its in-house model Mythos, it was able to port Claude to AMD GPUs in just three days — with almost no human involvement. If that’s true, it’s a pretty direct attack on Nvidia’s CUDA moat. That has been the big advantage until now: anyone who wanted to move AI models efficiently to alternative hardware needed a lot of specialized knowledge and time. For companies, this matters because hardware dependence directly affects costs, negotiating power, and scalability. Anthropic now wants to run Claude on four chip families to spread risk and reduce costs. So the real message is not “AMD is suddenly better,” but rather: AI is now also helping rebuild AI infrastructure itself. That’s almost a meta-plot. Source: The Decoder

🌏 US data brokers sell training data to China

An investigation by ChinaTalk suggests that US data brokers such as Mercor, Surge AI, and AfterQuery are selling training data worth around $500 million to Chinese labs — including RL environments originally developed for OpenAI or Anthropic. This is explosive for two reasons: First, it shows how valuable training data has become. Second, it reveals how hard AI export controls are to enforce in practice when data flows through intermediaries, service providers, and platforms. For the AI industry, this is a sign that competition is not decided by chips and models alone, but also by control over training infrastructure. And yes: once the same data circulates on global marketplaces, regulation quickly becomes a tracking exercise with very poor cards. Source: The Decoder

🛠️ Tool tip of the day

If you want to keep an eye on AI workflows, models, and automation, a tool for structured research, monitoring, and team notes is worth it. Especially for topics like benchmarks, agents, and enterprise AI, it saves a lot of time when making sense of the daily noise. Practical for anyone who doesn’t want to juggle ten tabs and three half-forgotten RSS feeds every morning. #


Want to avoid missing any news? Subscribe to the newsletter


Weekly AI news highlights

No spam. No ads. Just the essentials — concisely summarized. Weekly in your inbox.