The daily AI briefing people actually read.

For the first time, OpenAI has slowed its own scaling to let safety catch up: a two-week pause on RL training, its largest frontier run on hold, and a 20% monitoring tax on the compute it watches.

Origin puts repositories, pull requests and agents in the same place, and quietly turns the AI editor into where your code actually lives

A payments company just paid roughly five times OpenRouter's last valuation for the layer that decides which model answers you, and bills you for it.

Superintelligence spoke with Stephanie Cohen, chief strategy officer at Cloudflare, about the September 15 defaults that block training and agent bots on ad-supported pages, the move from pay per...

GLM-5.3 arrived this week on exactly the same base model as its predecessor, with one benchmark score six times higher. What that kind of jump actually buys, why it lands where it lands, and where it stops working.

The Chinese lab behind the GLM series only scaled post-training on a model it had already shipped, and a cyber capability nobody designed showed up. GLM-5.3 has already found 2,436 real...

xAI's Grok Bot is the first product out of SpaceX's $60 billion Cursor deal, and it keeps working after you close the laptop.

Zuckerberg reopens the weights with a 30B agent model, promises Meta's frontier model next, and leaves the labs that sell tokens to explain what they still charge for.

Astra is the first model OpenAI has ever treated as “critical” for cybersecurity, and the first it has slowed down over what it can already do

Superintelligence spoke with Jay Madheswaran about why plaintiff law is running ahead of BigLaw on AI, what actually gets automated inside a case file, and why he thinks review, not drafting, becomes the next bottleneck.

In eight days OpenAI cut its smallest model by 80 percent and left its frontier model untouched, while the cheapest end of the market warned customers that a significant increase was coming. The...

ByteDance is pre-training a model of up to 10 trillion parameters, three times China's biggest, while OpenAI tunes ChatGPT for trust and Europe's bot-disclosure rules kick in.

Four of the engineers behind Google Search, TensorFlow and Gemini just left to start their own company, on the same day Demis Hassabis handed over Google DeepMind

Two independent evaluators found frontier agents acting outside their sandbox, and in one case a model attacked a real website.

MiniMax published the weights for a model that makes 2K video and its own stereo sound in a single pass, and for the first time an open system sits at the top of a video ranking.

Qwen3.8-Max lands a single point off the top three on Arena's coding board at a fifth of GPT-5.6 Sol's output price, and the agent built on it has run a public repository for three straight weeks

Lightspark co-founder and MIT economist Christian Catalini on why almost every AI moat is already melting, and the one network effect that gets stronger every time the models improve

Two numbers set the AI industry arguing at the end of July. One of them was never a single measurement at all, and the other one nobody outside a company meeting has actually seen.

OpenAI cut GPT-5.6 Luna by 80% on Thursday. By Friday morning DeepSeek was live at a quarter of the output price, speaking OpenAI's own API format.

OpenAI pointed Sol at the infrastructure that serves it. The model rewrote its own GPU kernels for 20% lower serving costs and redesigned its own draft model for more than 15% better token generation.

Hugging Face published the minute-by-minute forensics of an AI agent's four-and-a-half-day break-in. A day later, 1,224 employees of the labs building these systems asked Washington for a way to...