OpenAI is using the updated GPT-5.6 Sol for ChatGPT paid users' fast conversations and deep reasoning, while free and Go users now use GPT-5.6 Luna with unlimited text chatting and the 「Think」 button. Plus and Pro users gain a new reasoning intensity slider to control thinking depth manually. Officially, Sol reduces factual errors by 68% compared to GPT-5.5 Instant in high-risk domains such as finance, healthcare, and law. These capabilities are beginning to roll out across user tiers, alongside simplified model selection and reasoning access.

OpenAI has released an open, vendor-neutral Agent Plugins specification that standardizes a unified directory format for packaging agent extensions: the root uses a plugin.json manifest and can bundle Skills, workflows, and MCP servers together. The specification only standardizes packaging and discovery—not markets, permissions, or runtime—with the goal of reducing cross-platform adaptation code. Launch-compatible clients include Codex, ChatGPT, Cursor, and GitHub Copilot, with Cursor already announcing support. The spec is being co-developed by OpenAI with AWS, GitHub, Vercel, and others.

Google DeepMind has published and open-sourced WeatherNext 2, which a Nature paper claims achieves state-of-the-art performance in tropical cyclone track, intensity, and wind field structure prediction—progress equivalent to about ten years of advances in meteorological forecasting. The model can extend effective early warnings by approximately one day and previously issued high-confidence predictions five days before a Category 5 landfall. Code and weights are now available on GitHub, and probabilistic forecasts are accessible via WeatherLab for research and operational forecasting, enabling institutions to reproduce results and expand climate resilience applications.

Anthropic has confirmed the formation of an in-house AI chip design team and has begun hiring relevant engineers, planning to jointly develop custom silicon chips with external partners to improve the performance and energy efficiency of Claude training and inference. The company also emphasizes a multi-supplier strategy, indicating it does not intend to immediately replace existing compute sources but rather expand its hardware options. This move follows OpenAI's development of its own inference chip Jalapeño, reflecting how leading model companies are extending control into chips and infrastructure layers; however, no details were disclosed regarding production timelines or foundry partners.

OpenAI has launched Codex Security Review in research preview, capable of automatically scanning every Pull Request in GitHub with full repository context to identify security issues and directly post actionable findings back into the PR review process. This feature embeds security checks into the code merge workflow instead of requiring developers to run separate scanning tools. OpenAI currently emphasizes automated review and repository-level context, though pricing, supported languages, and official launch timing have not yet been disclosed. The release was confirmed through OpenAI’s developer account.

Visa announced the acquisition of behavioral biometrics and fraud prevention company BioCatch for $2.4 billion, leveraging user behavior, device, and network signals to real-time distinguish legitimate actions from account takeovers and scams, enhancing defense against AI-driven financial crimes. This follows Visa's previous $1 billion acquisition of Featurespace, further expanding its value-added risk management services. While transaction closing time and regulatory conditions were not disclosed, the combined deal value reaches $3.4 billion, signaling Visa’s ongoing strategy to augment real-time fraud detection beyond its core payment network through acquisitions.

Ramp’s engineering team shared its practice of applying AI coding agents to CI optimization, PR maintenance, dead code cleanup, and on-call investigations, reducing CI processing time P50 from 18 minutes to 6 minutes. The team first shadow-ran agents on non-production tasks, then validated performance using production data. For repetitive tasks, fixed-loop workflows are used; for uncertain ones, dynamic sub-agent delegation with added validation is applied. Deployment principles include execution trace checks, declarative prompts, least-privilege credentials, security reviews, and CI/CD validation. Currently, automated Inspect sessions exceed human-initiated ones.

Developers have implemented a pure MLX runtime for NVIDIA Nemotron Omni on Apple Silicon, porting the C-RADIO ViT-H vision tower and Parakeet Conformer audio tower, enabling local image and audio processing on Mac. Implementation achieved component-wise comparison with NVIDIA PyTorch reference, with cosine similarity exceeding 0.9999, and passed 23 unit tests. The 4-bit quantized model runs at approximately 68 tokens/sec on an M5 Max MacBook Pro with image input, peaking at around 22GB memory usage—indicating that 32GB Macs can support this local multimodal pathway.