Welcome to the DailyAI Briefing, the day’s 5 largest AI tales. Amazon spent $42.4 billion additional cash than it introduced in over the previous yr, and it says AI is the explanation. A report claims two Chinese language labs educated their latest fashions on OpenAI’s GPT-6 Astra, Mistral has given away a trillion-parameter mannequin whereas Google retains its strongest one below lock, the free model of ChatGPT now flatters much less however will get extra details mistaken, and Reddit says Google’s AI Mode value it greater than 1 million visits every week.
Let’s dig in.
Amazon burns $42.4 billion on the AI build-out
The quantity is minus $42.4 billion. That was Amazon’s free money circulation for the 12 months to 30 September, in opposition to $14.8 billion left over a yr earlier, in accordance with its third-quarter results. Spending on property and gear rose 82% to $210.7 billion, which Amazon says “primarily reflects investments in artificial intelligence”, whereas money from operations grew a slower 29% to $168.3 billion.
The enterprise itself is ok. Gross sales rose 12% to $201.9 billion within the quarter and AWS, its cloud arm, grew 39%. However buyers have stopped giving the giants a free move. Amazon shares fell in early buying and selling on Friday, CNBC reported, a day after Alphabet’s document 13% drop, and Microsoft is now the one huge cloud firm nonetheless producing optimistic free money circulation, CNBC adds. Our studying, set out this afternoon, is that Large Tech is now paying for AI from reserves and borrowing slightly than spare money, which works solely so long as demand retains up. It follows our look at Google and Microsoft on Thursday.
Chinese language labs reportedly educated on GPT-6 Astra
DeepSeek and Moonshot AI “relied heavily” on outputs from OpenAI’s GPT-6 Astra to construct their newest fashions, The Information reported, citing individuals with direct information. The fashions named are Moonshot’s Kimi K3.1 Professional and DeepSeek’s V4.1-Flash 1017. Neither firm has addressed the declare, and no proof has been printed on how a lot Astra output was used.
The method is distillation, coaching one mannequin on one other’s solutions. It’s routine inside a lab and contested when performed to a rival with out permission. The report lands a month after OpenAI tied an earlier extraction campaign, with 16,000 tried requests from greater than 4,000 customers, to individuals linked to Moonshot. If it holds up, banning accounts after the actual fact isn’t stopping anybody, as we wrote this morning. For now the central declare rests on nameless sources, so look ahead to OpenAI to substantiate it.
Mistral provides away a trillion-parameter mannequin
Mistral has published the weights of Massive 4, nicknamed Le Chonk, so anybody with sufficient {hardware} can run it themselves. The model card lists about 1 trillion parameters, with round 52 billion lively for every phrase, a 1 million token context window and an Synthetic Evaluation Intelligence Index rating of 26. Mistral calls that the perfect of any Western open mannequin, whereas admitting the main Chinese language open fashions rating greater.
Want AI to save you hours a week? Our free playbook shows you how.Get the free guide →
At present, we launch the weights of a brand new and improved Le Chonk, advancing the open frontier for agentic coding and cybersecurity.
— Mistral AI October 29, 2026
The fascinating half is the distinction. Mistral pitches Massive 4 at cyber defenders, on the logic that an open mannequin can’t be lower off. Google went the opposite method with Gemini 4 Carbon, which scores 59 on the identical index and goes first to a small group of vetted defenders. Each arguments are cheap, which is why this break up is worth following: a mannequin good at discovering holes in software program is sweet at utilizing them too.
Free ChatGPT flatters much less and errs extra
From 29 October, the free model of ChatGPT provides shorter solutions with much less flattery, according to OpenAI’s release notes. The change covers Free and Go customers and anybody not signed in, whereas Plus and Professional solely get it on the backup mannequin. On OpenAI’s HealthBench take a look at, common solutions fell about 12%, from 3,361 to 2,970 characters.
The catch is within the system card addendum. OpenAI says the brand new model reveals greater factual-error charges than the 7 October one on all six measures it reported, although it does higher on teen security exams. Much less sycophancy is an actual repair, and it’s an trustworthy disclosure, however the sensible recommendation free of charge customers is apparent: count on blunter solutions and check anything that matters.
Reddit says AI Mode value it 1 million visits every week
Reddit’s visits from Google search fell by greater than 1 million every week on common in October, its letter to shareholders says, and it warns every day guests may drop by a low-double-digit share this quarter from about 130 million. “The reality is, AI Mode does not drive traffic,” chief govt Steve Huffman mentioned, as quoted by The Verge.
Though Google’s AI-powered modes cut back visitors to Reddit, Steve Huffman says that Reddit continues to be a ‘top search term’ on Google.
— The Verge (@verge) October 29, 2026
Income nonetheless rose 53% to $895 million, however the shares fell greater than 15%. Reddit’s reply is a login wall, examined first in some markets exterior the US. By itself numbers that’s logical, as we argued this morning, since logged-in customers herald many of the cash. It’s additionally a warning for each website that depends on Google, which is spreading ads in AI Mode slightly than pulling again.
Signing off
In order that was Friday, and the week ends the place it started: AI is costing greater than anybody is incomes from it but, and everyone seems to be combating over who will get to make use of whose work. Elsewhere on the positioning right now, Canada bought into Cohere with up to C$1.6 billion of public money.
Have a great weekend, and see you tomorrow.

