StationX
MEMBERSHIP

Sec & AI News — 13 February 2026

8 min readBy Nathan House
Get every new Sec & AI News issue
Straight to your inbox. No spam.

🎬 China's Video Models Make Everyone Else Look Polite

ByteDance's Seedance 2.0 landed. Four input modalities, 20-second clips, dual-channel audio, lip-sync good enough to earn an MPA denouncement inside 24 hours. Copyright enforcement: absent. Meanwhile Kuaishou's Kling 3.0 shipped 4K/60fps, native audio, six shots per clip — and generation times measured in geological epochs, until Leonardo AI hosted it and dropped that to minutes.

🇨🇳 China's Open-Weights Labs Undercut Everyone

Three drops, one theme: parity at a fraction of the price. Zhipu's GLM-5 — 744B params, MIT license, trained without a single Nvidia chip — beats GPT-5.2 and Opus 4.5 on multiple benchmarks at a sixth of the cost. Someone had it build a Game Boy Advance emulator autonomously over 24 hours. MiniMax M2.5 hits SWE-Bench parity with Opus 4.6 for a dollar an hour. And Alibaba's Qwen-Image-2.0 topped the text-to-image ELO leaderboard in a single 7B model. Intelligence too cheap to meter, delivered by the people you were told couldn't compete.

🧠 Gemini 3 Deep Think Solves Problems Humans Couldn't

Google's new reasoning mode is crushing ARC-AGI-2 at 84.6% and scoring 48.4% on Humanity's Last Exam with no tools. It solved 18 previously unsolved research problems and disproved a decade-old mathematical conjecture. Gold medals in physics and chemistry. Available exclusively to $250/month AI Ultra subscribers — Google's polite way of saying you can't test this anyway.

⚡ OpenAI's First Non-Nvidia Model Is Absurdly Fast

GPT-5.3 Codex Spark runs on Cerebras wafer-scale chips. 1,000+ tokens per second, 15x faster than standard Codex, and it built a playable Vampire Survivors clone in 50 seconds. OpenAI's first production deployment off Nvidia hardware, backed by a $10B+ Cerebras deal. Pro plan only, $200/month. The naming convention continues its downward spiral.

📢 ChatGPT Gets Ads, Loses Four Models, and Denies a Fake Orb

Busy week for OpenAI's PR team. Ads are now testing on Free and Go tiers — "clearly labeled," "don't influence answers," your chats stay "private from advertisers." Every ad platform said exactly that at launch. Same day, GPT-4o, GPT-4.1, GPT-4.1 Mini, and o4-Mini got deprecated; the 0.1% who still used 4o are very loud on Twitter. And that viral Skarsgard-fondling-a-silver-orb ad branded "Dime"? Not OpenAI. The real Jony Ive device slipped to 2027 and lost the "io" name to a lawsuit.

🎨 Generative AI Eats the Ad and Social Feed

The ad-industrial complex got its upgrade. Veo 3 is now in Google Ads Asset Studio next to Nano Banana — upload a product photo, get 15 video versions tuned per YouTube placement, and reportedly 22% lower CPA. Meta went the other way with gimmicks: animated Facebook profile pics via Meta AI, and Threads' "Dear Algo," where you literally type your feed preferences into a post. Born from a meme, shipped as a feature.

🖥️ Claude Cowork Hits Windows, Video Tools Go Node-Based

Anthropic's desktop agent — file management, multi-step tasks, MCP connectors — landed on Windows with full feature parity, Mac-only since January. Requires Claude Pro. Microsoft is quietly favoring Anthropic in 365 and Copilot, with Opus 4.6 now in Foundry. Elsewhere, Runway's "Story Panels" and Crea AI's "Prompt to Workflow" both shipped node-based builders — the simple text-to-video era is graduating into power-user territory.

🚕 Waymo's 6th-Gen Robotaxi: Fewer Sensors, More Cities

New van-shaped vehicle, 42% fewer sensors — 13 cameras (down from 29), 4 lidar, 6 radar, a fresh 17-megapixel imager. It handles rain and snow, on cheaper parts. Employee rides in SF and LA now, public access later this year. 2026 expansion targets Dallas, Denver, Detroit, Houston, Las Vegas, Nashville, Orlando, San Antonio, San Diego, Washington, and London.