Anthropic Launches Claude Platform on AWS in General Availability
AWS and Anthropic have announced the general availability (GA) of Claude Platform on AWS. This new service lets users access Anthropic's native Claude Platform experience directly through an AWS account — no separate credentials, contracts, or billing relationships required.
AWS becomes the first cloud provider to offer the native Claude Platform experience. The same APIs and features available directly from Anthropic are included, such as the Messages API, Claude Managed Agents (beta), advisor tools (beta), web search and fetch, and MCP connectors (beta).
This move signals that cloud integration of AI platforms is deepening further. For enterprises, the barrier to adoption drops significantly, since Claude can be used seamlessly on top of existing AWS infrastructure. The combination of AWS's global footprint and Anthropic's cutting-edge models could become a strong alternative to the Azure + OpenAI relationship.
Baidu Ernie 5.1 Competes with Top Models at 94% Lower Pretraining Cost
Baidu has announced its latest chatbot, Ernie 5.1. Despite having one-third the parameters of the previous generation, it achieves remarkable efficiency, requiring only 6% of the pretraining cost of comparable models.
The key to this cost reduction is an approach called "Once-For-All." It extracts smaller submodels from a single training run, eliminating the need for multiple full training runs. On the Search Arena leaderboard, Ernie 5.1 ranks 4th worldwide. The top three spots went to two Claude Opus variants and GPT-5.5 Search, making this a case of a Chinese-made model breaking into the top tier.
Training costs for AI models have become a multi-billion-dollar challenge for the industry as a whole. If Baidu's method can be applied by others, it has the potential to change the very economics of AI development. That said, how much of what relies on their proprietary architecture can be generalized remains to be seen in detailed future evaluations.
EU Regulators Face Divergent Responses from OpenAI and Anthropic
According to The Decoder, a clear gap has emerged between OpenAI and Anthropic in their responses to the European Commission's AI regulatory efforts.
OpenAI has offered the European Commission direct access for a security review of its new model, GPT-5.5 Cyber. Anthropic, on the other hand, still has not provided access to regulators, even after four to five rounds of talks concerning its Mythos model.
This situation highlights just how much EU AI oversight depends on the voluntary cooperation of the companies being regulated. With the AI Act coming into force, how to secure actual model access remains an institutional challenge. The contrast between OpenAI's cooperative stance and Anthropic's cautious response can also be read as reflecting each company's strategic positioning on regulation.
Hackers Abuse Claude.ai Chats to Distribute Mac Malware
According to BleepingComputer, an attack has been observed in which hackers combined Google ads with Claude.ai's chat feature to distribute malware to Mac users.
This technique is a real-world example of AI chat tools being exploited as instruments of cyberattacks. An attack vector in which malicious content is delivered through conversations on AI platforms that users trust is difficult to anticipate with conventional security measures.
As AI tools proliferate, attacks abusing these platforms are expected to increase. Providers will need to strengthen safeguards to protect users, and users themselves should stay alert to files and links distributed through AI-based conversations.
Nobel Economics Laureate Acemoglu Offers a Sober View on AI Productivity
In an interview with MIT Technology Review, Daron Acemoglu, winner of the 2024 Nobel Prize in Economics, shared his thoughts on AI.
Although his views have been unpopular in Silicon Valley, two years ago Acemoglu estimated that AI would give US productivity only a "modest boost" and would not eliminate the need for human work. Two years on, his tempered outlook is still not widely accepted. Anxieties about AI bringing an end to human jobs are being voiced everywhere.
Nevertheless, Acemoglu maintains his fundamental position: "AI is reasonably good at automating specific tasks, but quite a few jobs will be perfectly fine." His observation that a gap persists between the "complete transformation of white-collar work" promised by big tech CEOs and real-world productivity gains offers an important counterweight amid swirling, bubble-like expectations.
MiniCPM-V 4.6: An Ultra-Compact Multimodal LLM That Runs on Smartphones
MiniCPM-V 4.6, which has been trending on HuggingFace, is an ultra-compact multimodal LLM built on SigLIP2-400M and Qwen3.5-0.8B. Optimized for edge deployment, it is reportedly capable of image and video understanding even on smartphones.
Its key features include inheriting single-image, multi-image, and video understanding capabilities while dramatically improving computational efficiency. With the introduction of 4x/16x selective visual token compression, it stands as the model in the MiniCPM-V family best suited for edge deployment.
Meanwhile, in the LocalLLaMA community on Reddit, a demo also drew attention in which Gemma 4 runs fully offline with Transformers.js on WebGPU and controls the robot "Reachy Mini" over WebSerial. This proof that an LLM can run in the browser and operate a robot is a symbolic milestone showing how far edge AI has come.