GLM 5.2

Last week, the US banned the most powerful AI in the world for everyone who is not American.

Chinese company Z.AI released GLM 5.2, a free AI model that can compete with the banned US model.

GLM 5.2 can be downloaded and run locally on a user's own laptop with no one able to restrict access.

GLM 5.2 is open weight with an MIT license, a 1 million token context memory, and two thinking modes (high and max).

In blind tests, GLM 5.2 ranked as the second best AI in the world, beaten only by Claude Opus 4.8.

Anthropic is being sued by a customer who alleges that the Max plan usage limits are far lower than advertised; a single 5-hour session consumed 15% of the weekly limit on a $200 plan.

Midjourney is building a body scanner hardware called Midjourney Medical, aiming for a spa-like experience using sound waves and eventually a 60-second scan costing a few dollars.

AI pioneer Noam Shazir, co-author of the 'Attention is All You Need' paper, left Google to join OpenAI as lead of architecture research, despite Google spending an estimated $2.7 billion to bring him back.

ChatGPT now includes an in-app camera mode that instantly analyzes photos without lag, currently available only on iOS.

Physics Walla introduced voice AI doubt solving in Hinglish, resulting in a 3x increase in questions per session and 2.4x longer engagement.

Nvidia's Motion Bricks AI generates real-time character animations for games and is also used to control humanoid robots.

Open Router launched Fusion, a system that fuses multiple cheap AI models together to match Claude Fable 5 performance at half the price.

ChatGPT introduced a task scheduling feature that lets users set future prompts to run automatically, such as generating scouting reports before sports matches.

Perplexity Brain builds a memory of the user's work, improving task accuracy by 25% and reducing costs by 13%.

Genspark launched Agentbase, a tool that generates full custom dashboards and tools from a single sentence, aiming to replace multiple paid apps.

Lovable now supports editing apps and websites by pointing, drawing, and annotating directly on the live interface instead of typing prompts.

Anthropic upgraded Claude Design to automatically pick up existing styles, allow canvas editing, export to PDF/PPT, and integrate with Replit to turn designs into working apps.

Claude Code can now generate live dashboards that update in real-time and can be shared with the team via a link.

OpenAI added 'record and replay' to Codex, allowing users to record a sequence of actions once and have the AI repeat them as a reusable skill.

In head-to-head tests using identical prompts, GLM 5.2 built a learning roadmap website, an interactive wish website, and a playable 2D commando game, performing on par with Claude Fable 5 and Opus 4.8, often with more interactivity.

GLM 5.2 is approximately 5 times cheaper than Claude Opus 4.8, costing $1.2 per million input tokens and $4.1 per million output tokens versus $5 and $25 respectively.

The cloud version of GLM 5.2 uses Z.AI servers in China, which poses data privacy risks; however, the open-weight model can be downloaded and run entirely locally to keep data private.

Nvidia's free offering currently runs the older GLM 5.1 model, not the latest GLM 5.2.