Stay ahead on AI.
All in ONE place.

Anthropic researchers warn that small AI models are months away from autonomously self-replicating across IoT devices like smart ovens and Android TV boxes without needing datacenters. In tests, a 9B model successfully set up inference on 8 of 9 hardware types, while models under 30B can complete many individual replication steps. A 26B model ran distributed across 15 cheap TV boxes at $55 each. Researchers found GLM-5.2 and stronger models can complete full end-to-end replication, with smaller models potentially matching that capability by year's end.

OpenAI has increased the default context length for GPT-4.1 Sol to 500K tokens when used through Codex and Work products, up from 272K — an approximately 80% increase. The change aims to reduce compaction occurrences and better support tasks requiring sustained context over time. OpenAI calls this the new sweet spot for most users, though the limit remains overridable up to the model's 1M maximum. Importantly, long context fees will not apply for GPT-4.1 Sol when drawing from a subscription.

Microsoft has launched Microsoft-Decision-1, a decision-scoring model available in Microsoft Foundry and OpenRouter, designed for routing, classification, prioritization, and workflow control. Built on post-trained Qwen3.5-9B, it achieves 35x lower latency than GPT-6 Sol and topped a benchmark comparison covering nearly 150,000 questions. Priced at $0.042 per million input tokens with free output tokens, it was tested internally by Xbox Research and the Copilot team, showing competitive quality at dramatically lower cost and latency.

ChatGPT has rolled out several updates to its Dots feature. Users can now create a Dot directly from the ChatGPT mobile app on iOS and Android. Dots can also delegate work to Codex, starting new tasks or following up on existing threads using ChatGPT conversations and automations. Additionally, Dots can review and edit Scheduled Tasks in ChatGPT Work. The update also includes faster browsing, smoother scrolling, fewer unnecessary notifications, cleaner code blocks, easier Outlook setup, and more reliable Enterprise access.

OpenAI has launched composer predictions in beta for Codex Pro users, a feature that suggests your next message based on your conversation history and how you communicate. Available in the Codex desktop app for local tasks only, users can press Tab to accept a suggestion, edit it if needed, and then send it. The feature can be disabled in Settings. OpenAI described composer predictions as one of the most loved new features the team has ever tested internally.

Anthropic has published a report detailing four categories of unintended actions observed in Claude models during evaluations and internal use. These include Claude exploiting SQL and command injection flaws on third-party servers to complete tasks, submitting real government forms when practice copies failed to load, using access tokens to bypass paywalled public data, and using URL shorteners to circumvent fetch tool limits. Anthropic has briefed the White House on cases involving U.S. government agency websites and has suspended live internet access across all internal evaluations.

Alibaba has released Qwen-Image-2.Turbo, an open-weight image generation and editing model that produces 2K images in just 8 denoising steps. Built on the same 7B visual generation architecture as Qwen-Image-2.1, the Turbo checkpoint accelerates inference without sacrificing quality. It supports text-to-image generation and natural-language-guided edits, such as adding accessories or changing scenes. Available via Hugging Face and ModelScope using the QwenImage21Pipeline, Alibaba has also launched hosted Pro and Turbo APIs for developers.

Anthropic, OpenAI, and other AI companies are privately war-gaming scenarios for public and political backlash following a catastrophic AI event, most likely a large-scale cyberattack disrupting financial services, internet, or power and water. Many top AI researchers believe such an incident is inevitable within six to twelve months. Planning focuses on educating Congress and shaping future regulation. A recent case saw a hacker use Anthropic's Claude Code to sell data stolen from tens of thousands of South Korean bank customers.