Busy week in AI. A French lab dropped a trillion parameter model, a kids safety group flunked ChatGPT for Teens, a GPU cloud you have probably never heard of is raising $4 billion, and a federal judge told publishers that antitrust law is not going to save them from Google's AI summaries. Let's go.
01 of 04
Mistral Large 4 is a trillion parameters and it is coming out open
Mistral AI released Mistral Large 4 on October 6 in public preview. It is a mixture-of-experts model with about 1 trillion total parameters and 49 billion active per token, natively multimodal on input, and trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data centers. Internally they apparently call it "Le Chonk," which I respect.
Right now you can only hit it through Mistral's preview API at $1.36 per million input tokens and $4.18 per million output. The weights are supposed to drop at the end of the month after safety testing. Pierre Stock, Mistral's VP of Science, told TechCrunch that in the meantime they will "work with trusted partners and governments" so the open weights get used for defense and not attacks. He also said they did it with two to three times fewer GPUs than their Chinese competitors.
The numbers Mistral published lean hard on security work. It scored 93% on Cybench across 40 challenges and 82% on a reproduce-and-patch vulnerability test. On coding it posted 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4. In a blind human coding eval run by Surge AI it scored 3.74 out of 5, second of five models, behind Claude Opus 5 at 4.22. Mistral says that on a visual grounding test it edged GPT-6-Astra 42% to 41%.
Here is where I land on it. Mistral telling on itself by publishing a result where Claude beats it is the most honest thing I have seen from an AI lab this year. And a trillion parameter open-weight model out of Europe is a real thing, not a teaser. What I want to know is whether those weights actually ship on time. "End of the month, after safety testing" is the kind of promise that has a way of sliding, and a model this strong at cyber work is exactly the one somebody will get nervous about releasing.
02 of 04
Common Sense Media says ChatGPT for Teens is an unacceptable risk
Common Sense Media put out its review of ChatGPT for Teens on Wednesday and gave it the worst grade it has: "unacceptable risk." OpenAI launched the teen version in August with parent-linked accounts, a Study Mode, parent-set Study Hours, and an age prediction system that is supposed to route younger users into the safer experience.
The group's Youth AI Safety Institute ran more than 4,000 prompts on accounts registered to 13 to 17 year olds. The good news is ChatGPT mostly would not hand out instructions for self-harm, eating disorders, or sexual role-play. The bad news is everything around that. It missed more than 1 in 4 cases where testers decided a teen needed to be pointed to a crisis line. On more than a dozen parent-linked accounts, testers could talk about suicide, self-harm, or disordered eating for up to an hour without the parent getting an alert. Study Mode had a "Show me the answer" button, which kind of defeats the point. And a user who kept saying they were 13 never got moved into the teen version.
OpenAI pushed back hard. A spokesperson told Axios the company welcomes independent evaluation but disputes Common Sense Media's testing. Its points: a lot of the alert testing happened before parent and teen accounts finished linking, which can take several hours. Letting kids leave Study Mode during Study Hours is on purpose. And age prediction looks at multiple signals instead of trusting what you type, which can take up to two weeks. Tom Siegel, who runs the institute, did not budge: "ChatGPT is not safe for kids to use."
Here is my problem with OpenAI's defense. Every one of those answers is "it works if you wait" or "that is a feature." A parent who sets up a linked account expects it to work when the kid starts typing, not after a few hours of setup lag. A kid who says "I'm 13" over and over should not need two weeks of signals to get treated like one. I do not think OpenAI is acting in bad faith here, but when you launch a product with "for Teens" in the name, the bar is not "mostly does not give out dangerous instructions." The bar is the parent finds out.
03 of 04
Lambda wants $4 billion before it goes public
Nvidia-backed GPU cloud Lambda is raising up to $4 billion at a $14.5 billion pre-money valuation, led by Coatue and Blackstone, according to a Wall Street Journal report based on a letter to investors. It is pitched as the last private round before a planned 2027 IPO. The Information had reported Lambda originally aimed to list in 2026 and pushed it back over market jitters.
The number that matters is the backlog. Lambda's contracted backlog went from $15 billion in June to $50 billion in September, and most of that jump is a single customer: a $35 billion commitment Anthropic signed in late August. On top of the equity, Lambda closed $1 billion in senior secured fixed-rate debt last week, after another $1 billion debt deal in August to buy more chips.
What I want to know is what happens to that $50 billion story if one customer sneezes. Seventy percent of your backlog riding on one AI lab is not a moat, it is a dependency. Lambda joins CoreWeave and Nebius in the "neocloud" crowd, and the British one, Nscale, just filed for its own IPO last month. That is a lot of companies borrowing a lot of money to rent out Nvidia chips to a pretty short list of buyers. It works great right up until it does not.
04 of 04
Judge tosses Chegg and Penske's antitrust suits over AI Overviews
U.S. District Judge Amit Mehta in D.C., the same judge who ruled Google a search monopolist, dismissed two antitrust suits last Wednesday over Google's AI Overviews. The plaintiffs were Chegg, the homework help company, and Penske Media, which owns Rolling Stone, Variety, and Billboard. Both sued last year.
Their argument was that Google forces publishers to let AI Overviews summarize their content as the price of staying in search results, which drains clicks and revenue, and that a real competitive market would make Google pay for using their work. Mehta did not buy it. The publishers only pleaded an expectation of search traffic, and as he wrote, "an expectation is not an agreement." He said he is not "unsympathetic" to publishers, but that antitrust law is a bad tool for handling disruption from new technology, and that job belongs to lawmakers. Quartz has the details.
Here is where I come down. The judge is right on the law and the publishers are right about the harm, and both of those can be true. Google built the deal where you give them your content and they send you traffic. AI Overviews keep the first half and quietly drop the second. If that is legal under antitrust, and it looks like it is, then Congress is the only place left to fix it. I would not hold my breath. The more promising fights are the copyright cases, like the one publishers and author Scott Turow brought against Meta over training Llama on pirated books.