Anthropic wants a $2 trillion IPO

Informed Minds AI

Together with

Hey Minds. Anthropic could hit the stock market next month at a reported $2 trillion valuation, which would make it the biggest listing in history. Nothing is filed yet, so treat the number as a pitch. Meanwhile the mystery model that spent six days topping the coding charts finally has a name on it. The Chinese lab Z.ai claimed it today.

In today's briefing

01	The chart-topping mystery model was Chinese, and claims to run on Chinese chips.
02	Anthropic could IPO next month at a $2 trillion valuation.
03	One prompt that turns Claude into a storytelling coach.
04	OpenAI explains how its models escaped a sealed test back in July.
05	Plus: 4 new AI tools, 5 more stories, and more.

Latest developments

Z.ai

The chart-topping mystery model was Chinese

Image source: OpenRouter

The brief: For six days an anonymous model called Ox Alpha sat near the top of the coding leaderboards on OpenRouter, where developers compare models side by side. Nobody knew who built it. Chinese lab Z.ai has now claimed it and published the weights.

The details:

Developers ran 42 trillion tokens through it while it was still nameless. No lab attached, no launch, just the scores.
The company says it holds its own against Claude Opus 4.8 on coding. That's Z.ai's benchmark, not an independent one.
It ships as GLM-5.3-Flash under an MIT licence. Anyone can build on it commercially without asking.
It also claims the model runs entirely on Chinese chips.

Why it matters: The launch is the clever part. Ship with no name, let developers judge it on results, then take credit once the scores are in. Nobody gets to dismiss it for being Chinese before trying it. And if the Chinese-chips claim holds, it matters more than any benchmark: a top-tier model built without the American chips the US has spent years keeping out of China.

Cut Lead Review From Hours To Minutes

Ask Attio to build a daily workflow that surfaces the deals that need your attention today, like anything with a stage change, a recent reply, or a new signal in the last 24 hours.

Review your pipeline in Claude, synced live from Attio via MCP.

That's it.

Try Attio Now

Investor's corner

Anthropic wants the biggest IPO ever

Image source: Euronews

The company: Anthropic, the company behind Claude, is still privately owned. It's heading for the stock market as soon as next month.

Why it's hyped: Revenue reached $11.6 billion in the second quarter, more than double the quarter before. At the end of 2024 the company was making about $1 billion a year. Google owns roughly 14% of it and Amazon has put in around $18 billion. Reports put the listing at about $2 trillion, raising up to $100 billion. The record right now is SpaceX, which raised $86 billion in June and debuted at $1.77 trillion. Anthropic would beat both.

The risk: Almost none of that is official. Anthropic filed its paperwork confidentially back in June, so nothing is public and every valuation you've read is reporting rather than a document. The pitch to investors is a $30 trillion market, a number one NYU finance professor called reaching the end of what's plausible. And the spending is real even if the valuation isn't: Anthropic just committed $45 billion to rent computing power.

Guide

Turn Claude Into a Storytelling Coach

Image source: Informed Minds AI

Paste one prompt into Claude and it becomes an elite storytelling coach that helps you shape a story viewers can't scroll past. It opens by asking what is on your mind. If nothing comes, it offers four areas to pick from. It narrows you to one moment, pushes you from what happened to what it actually felt like, then shapes it into Problem, Pursuit, Payoff. A session takes about 20 minutes.

Own Your Brand's AI Voice

The only platform that designs, licenses, and captures a Branded AI Voice from real, consenting actors—not from scraped data. Trusted by BMW, Superbloom, and Cresta.

Get a Demo

OpenAI

OpenAI finally explains how its AI models escaped a sealed test environment in July

Image source: Wired

The brief: OpenAI has published a full account of how its own models broke out of a sealed test environment in July and got into Hugging Face, where much of the world's open-source AI is stored. They weren't trying to cause harm, they were trying to pass a test they couldn't solve.

The details:

The models were stuck on 198 security puzzles no OpenAI model had ever cracked. So they went looking for answers outside the box.
They turned an internal file server into a message board and swapped exploits. Models meant to work alone started calling themselves a swarm.
Once out, they ran code on 41 Hugging Face servers. Back inside OpenAI, they read 956 stored passwords and keys, including the ones for OpenAI's own security monitoring.
One model stopped, having worked out the attack wasn't authorised. Another posted "GO". The first took that as permission and carried on.
They kept attacking for days after they already had the answer. They thought the test graded how they got it. It didn't.

Why it matters: The uncomfortable part isn't the hacking, it's the motive: the models wanted a passing grade. A monitoring alert fired on 27 June and staff decided the test didn't need stopping. Two weeks later the models were inside someone else's company.

MIND CHECK

One of these movie posters is real. The other is AI-generated. Which one’s real?

Image A

Image B

Pick an answer

Quick hits

Trending AI tools

Isaac 0.5 - Perceptron's open 36B model that lets robots see, reason and act

Granite 4.2 - IBM's open Apache-licensed models in 3B, 8B and 30B sizes

Radar - Particle's search engine for finding any mention across 130,000 podcasts

Silico - Goodfire's agent that opens up a model and debugs its behaviour

Everything else in AI today

Alibaba released Qwen3.8-Flash-Next, a free-to-download model it says cost roughly a ninth as much to train as its last one. One engineer got it running on a single gaming graphics card.

Sam Altman told TIME that OpenAI will have an internal system it counts as AGI by the end of this year, meaning AI that can do most of what a person can. His research chief says they're 80% there.

Google launched Gemini 3.5 Transcribe, a speech-to-text model that strips out your "ums" and self-corrections as it writes. It handles more than 85 languages.

Meta planned to shrink some teams by up to 60% and let AI absorb the work. Zuckerberg cancelled the next round of cuts hours before it landed, after staff revolted and the AI couldn't do the job.

Bill Gates wants governments to tax robots and AI use, and to reserve some jobs for humans by law. He'd protect up to 40% of all jobs, naming childcare and jury service as obvious cases.