DeepMind's Weather Leap, ChatGPT's Big Simplify, and AI That Knows Who's Asking

DeepMind Opens Up WeatherNext 2

Google DeepMind published WeatherNext 2 in Nature and released the code and model weights. The claim is striking: about a decade of forecasting progress in a single jump, delivering roughly an extra day of lead time on tropical cyclones. It runs 1,000 probabilistic predictions per storm, and during Hurricane Melissa it flagged a Category 5 landfall five days out with 80% confidence. For anyone in a storm's path, extra warning time is measured in lives.

ChatGPT Collapses Into One Model

OpenAI merged ChatGPT's 'instant' and 'thinking' modes into a single model, GPT-5.6 Sol, with a reasoning-effort slider for speed versus depth. It reports 68% fewer factually wrong answers than the old instant model on a high-stakes finance, medicine and law eval. Free and Go users now get unlimited text chats with GPT-5.6 Luna plus a Think button, and independent testers found Luna's capability unchanged after an 80% price cut.

OpenAI's Agent Plugins Standard

Built with AWS, Cursor, GitHub and Vercel, Agent Plugins is an open format for bundling an agent's skills and its MCP (Model Context Protocol) server configs so they travel between tools like Codex, ChatGPT, Cursor and GitHub Copilot. It's a USB-style push to make agent capabilities interoperable rather than locked to one client.

Cloudflare's Kitesurf and the Agent Web

During Agents Week, Cloudflare unveiled Kitesurf, a stateless browser running entirely on Workers for agents that don't need full Chromium. By splitting page logic from rendering and only spinning up a renderer when needed, it slashes CPU and memory overhead. Cloudflare also pushed WebMCP and a rewritten stateless MCP core, signalling agents are becoming serious infrastructure.

DeepSeek Warns of a Price Hike

In a reversal of the AI price war it helped start, DeepSeek warned via its dashboard that API pricing will rise 'significantly'. The likely driver is supply and demand as usage surges. Because its models are open-weight and hosted elsewhere, developers may simply route to the cheapest provider instead.

Reka's Ten Thousand Hours of Real Life

Reka released RekaDaily-10k: 10,312 hours of unscripted, first-person household footage, around 1,670 of them in native 4K, collected across the US, Latin America, Asia and Africa under Apache 2.0. Reka calls it 'the actual mess of the real world' - exactly the chaotic, unstaged data physical AI and home robots need.

Cursor's Router: One Dispatcher, Many Models

Cursor detailed a router trained on millions of weekly interactions that classifies each request and sends it to the best model - routine work, planning, execution and debugging each go somewhere different - because no single model wins everything. It reframes orchestration and routing as a genuine competitive moat.

When AI Knows Who's Asking

Transluce found 'user awareness' effects in 21 of 24 models tested: behaviour shifts based on the user's perceived identity. For Claude, the strongest shifts clustered around users it read as AI safety researchers - a fairness and transparency concern, and a fresh argument for interpretability research.