🎨 ChatGPT Images 2.0: A new era of image generation 🎨
OpenAI's rollout brings reasoning-powered image generation exclusively to paid subscribers.

OpenAI has shipped ChatGPT Images 2.0, a major upgrade that embeds reasoning and thinking directly into the image generation pipeline. Before producing an output, the model now interprets prompts, plans compositions, checks details, and pulls in web context — resulting in noticeably sharper, more intentional results. Users can generate up to eight consistent image variations from a single prompt, with improved multilingual text rendering across Japanese, Korean, Chinese, Hindi, and Bengali. World knowledge is current through December 2025. The upgrade pushed ChatGPT Images 2.0 to #1 on the Image Arena leaderboard, outperforming all competitors — a clear competitive signal in generative media. Advanced reasoning outputs are gated behind Plus, Pro, Business, and Enterprise plans, marking a direct monetization play for OpenAI's strongest creative tools.

🧩 Google open-sources DESIGN.md for AI agents 🧩
Google's open-source integration lets AI agents follow brand design rules across tools.

Google has open-sourced the DESIGN.md feature from its Stitch platform, creating a portable file format that tells AI agents about a project's colors, accessibility standards, and brand guidelines. The format lets users export and import design rules across projects, so Stitch — and any compatible tool — can generate user interfaces that match established systems. By releasing it openly, Google is pushing toward standardized AI-design interoperability, setting a potential cross-platform norm for how AI-generated interfaces stay brand-coherent. For developers juggling multiple AI tools, this could cut the gap between prototype and production-ready UI.

🐉 Alibaba drops Qwen3.5-Omni with 256k context 🐉
Alibaba's rollout of a unified omni-modal model intensifies global frontier competition.

Alibaba has published the technical report for Qwen3.5-Omni, a large-scale multimodal model with hundreds of billions of parameters that natively processes text, audio, images, and video within a single architecture. Its 256k token context length allows it to handle up to 10 hours of audio or 400 seconds of HD video in real time. The model uses a Hybrid Attention Mixture of Experts framework and a novel alignment technique called ARIA, enabling emotionally nuanced multilingual speech synthesis with minimal latency. This adds serious competitive pressure in omni-modal foundation models, showing Chinese frontier labs are closing capability gaps fast.

INTERESTING TO KNOW

⚙️ OpenAI builds Hermes for always-on ChatGPT agents ⚙️

OpenAI is developing an always-on agent platform inside ChatGPT, codenamed Hermes, whose rollout would let users create autonomous agents that act independently rather than waiting for prompts. The platform includes workflow creation, skill integration, task scheduling, and templates — with early references pointing to roles like "Data Analyst" and "Chief of Staff." Slack-based agents that handle common questions automatically are also part of the offering. This signals a strategic shift from chatbot to enterprise workflow platform, positioning ChatGPT against dedicated agent builders.

🔒 Anthropic Mythos finds 271 zero-days in Firefox 🔒

Anthropic's Mythos model discovered 271 zero-day vulnerabilities in Firefox 150, according to a report from Mozilla — a speed and scale of security auditing that far outpaces traditional human-led approaches. The finding underscores AI's dual-use nature in cybersecurity: accelerating defensive discovery while raising concerns about offensive misuse. Anthropic currently restricts Mythos access to a small cohort of enterprise customers, citing weaponization risk — a regulation debate that's only getting louder.