🎬 Gemini Omni 1.1 Flash ships 4K video upscaling 🎬
Google's latest rollout adds advanced video controls and 4K output via the Gemini API.

Google launched Gemini Omni 1.1 Flash, a major upgrade to its video-generation capabilities now available through the Gemini API. The model introduces new controls allowing developers to extend scenes up to 40 seconds, interpolate between first and last frames, and perform 4K upscaling for higher-resolution output. These features are built to enable faster video iteration for creative and production workflows. The upgraded model has also climbed to the top position on Arena’s text-to-image leaderboard, applying direct competitive pressure on rivals like Runway and Sora. The release marks Google’s intensifying push in multimodal generative media, challenging existing video-creation tools on both capability and speed.

🔧 Anthropic ships Model Hardware Standard for AI agents 🔧
A new open standard for hardware integration lets AI agents control lab equipment in minutes.

Anthropic unveiled the Model Hardware Standard (MHS) in research preview, a model-agnostic specification enabling AI agents to operate physical scientific and manufacturing equipment such as microscopes, robotic arms, and lab instruments. Machine owners describe their equipment in natural language, and MHS converts those descriptions into structured reference files any compatible agent can interpret. Setup times reportedly drop from weeks to hours or minutes. In a notable demo, Claude autonomously learned to align a laser through trial and error. Launch partners include Tecan, QIAGEN, AWS, Hugging Face, and Raspberry Pi, with an open-source release planned later. This expands AI agent deployment from software into physical-world operations, opening a new adoption frontier.

💻 Codex adds persistent reasoning to its SDK 💻
OpenAI refines Codex's integration layer for developers building on custom model backends.

OpenAI’s Codex coding agent received an update adding “persistent” as a recognized value in the reasoning-effort protocol and TypeScript SDK types. When a developer targets a custom Responses-compatible provider whose model specifies its effort level as persistent, the system now correctly deserializes it to the new Persistent variant instead of leaving it as a custom string. Functionally, the value is rewritten to “disabled,” ensuring backward compatibility. While technically narrow, the change matters for developers building on Codex with custom model backends and signals OpenAI’s commitment to developer-facing extensibility. The update was merged via a public pull request on GitHub, keeping Codex’s development transparent.

INTERESTING TO KNOW

📄 Cohere launches Parse at $1.50 per 1K pages 📄

Cohere launched Parse, a new enterprise document-intelligence product with pricing set at $1.50 per 1,000 pages via the Cohere API. Built on a cost-effective vision language model, Parse converts complex multimodal files — PDFs, scanned forms, mixed-format documents — into structured, machine-readable data across nine major languages. A free version is available through Cohere Space for experimentation. The tool targets a persistent enterprise bottleneck — extracting reliable structured data at scale — positioning Cohere competitively against larger players in AI infrastructure.

🧠 UK trials live AI guidance in brain surgery 🧠

In a milestone for real-time surgical AI integration, surgeons at University College London Hospitals performed the first brain surgery assisted by a live AI system. The tool watched the operation through a surgical camera and flagged hidden arteries and optic nerves in real time, helping surgeons avoid critical structures during pituitary tumor removal. Patient Rhys Hibbert had his vision restored within days. The team is now moving toward larger clinical trials, a clear adoption signal for AI-enabled healthcare in the UK.