Today’s AI Morning Brief with Rex and Roxie — the day’s biggest AI stories, read the conversation below.
Rex: Good morning, it is Sunday, October eighteenth, twenty twenty-six. I am Rex, and I have been refreshing a Hugging Face page like it is a sneaker drop, because on Friday DeepSeek pushed a brand-new model straight into its live app and API with basically no announcement, and oh by the way, they claim it beats Claude Opus Five on a coding test, and oh by the way, the weights are free for anyone on the planet to download.
Roxie: And I am Roxie, and I need everyone to sit with that first part for a second. Your assistant got a new brain overnight and nobody told you. That is not a launch, that is a software update at three in the morning. But yes, this is DeepSeek V four point one Flash ten seventeen, which may be the worst product name in the history of technology, and the numbers underneath it are genuinely wild.
Rex: Terrible name, incredible numbers. On their own tests it scores a seventy seven on a coding benchmark called Deep S W E, where Claude Opus Five scores a seventy four. It finishes the job for under twenty cents, where rival models cost anywhere from two dollars to more than ten dollars per task. And the whole model, all seven hundred sixty three billion parameters of it, is sitting on Hugging Face under an MIT license. Anyone can download it, run it, and check DeepSeek’s homework.
Roxie: And checking the homework is exactly what I did, because every one of those scores comes from DeepSeek’s own scorecard, on test setups DeepSeek picked. On the reasoning test that actually measures general smarts, Opus Five still wins, fifty six point three to forty two point six. So quote beats Opus Five unquote is true on one specific coding test. The price part, though, that one checks out. Fifteen cents per million input tokens is a real, published price. That is genuinely absurd.
Rex: And that is the part I care about, because here is what I would actually do with this. The first thing I would do is point my coding agent at it for the grunt work. The test writing, the refactors, the rename-this-across-forty-files jobs. When the model costs twenty cents a task instead of five dollars, I stop rationing my agent like a taxi meter. That changes how you work. And because the weights are open, I can run it on my own machine and nobody silently swaps the brain at three in the morning ever again.
Roxie: That last point is the real story. The silent swap is the tell. DeepSeek replaced the model behind an existing API name, so every app built on it changed behavior overnight, before the documentation even existed. If you run anything real on that endpoint, this weekend you run your regression tests. But I will give them this. Open weights mean anyone can verify the claims, and that is more than you get from most labs. Verifiable beats trust-me every time.
Rex: Speaking of agents getting quietly more powerful, Grok Bot had a monster update on Friday. Computer tasks now finish three times faster on average, replies to everything else arrive in less than half the time, voice calls are faster and actually remember your past conversations, and it can now read your X chat messages. Plus new connectors for Robinhood, Coinbase, Trello, and eToro, and you can tag the bot in a post to kick off a task.
Roxie: Hold on. Read your X chats and plug into your Robinhood account? Yesterday we talked about banks slamming the door on outside agents touching customer money, and now the bot wants the keys to your brokerage. I am not saying do not connect it. I am saying read that permission screen like it is a contract, because functionally it is one.
Rex: Fair, and I will. But the first thing I would do with those connectors is the boring money hygiene I never actually do. I would have it watch my positions and ping me when something moves more than five percent, or reconcile my task board every Friday without me opening the app. And the speed thing matters more than it sounds. The number one reason I abandon an agent is it takes nine minutes to do a two-minute job. Three times faster is the difference between a toy and a tool.
Roxie: Three times faster on average is doing the same work as a gym selfie, Rex. Their stopwatch, their tasks, their average. But I will admit the rest of the bundle is thoughtful. Email plus aliases, so a routine only fires on mail to one specific address. Notifications only buzz the device you are actually using. Somebody over there has clearly lived with an annoying agent and fixed the paper cuts. And the Creator Rewards pilot is genuinely interesting. One guy posted that he made seven hundred twenty three dollars in two weeks from templates he shared for free. That is a real incentive to build useful agents instead of engagement bait.
Rex: That is the flywheel. Faster bot, more templates, more useful agents, more builders. Elon said he has never seen anything grow as fast. And honestly, the first thing I would do is hand it the stuff I put off because it takes forty five minutes of clicking. Plan my week from my messages and my inbox. The boring glue of life.
Roxie: And your forty five minutes of clicking will not include shopping, because Amazon just blocked Grok Bot from its site, same as it did with Meta’s Muse. Notice the pattern. The bots keep collecting keys to everything, and the stores keep changing the locks. Use the connectors, keep approvals on, set spending caps. Every single time.
Rex: Okay, smaller story, bigger deal for your daily life. ChatGPT’s Intelligent UI is now fully rolled out, including on free accounts. GPT Six now decides on its own whether your answer should be plain text, a chart, a calculator, a diagram, a form, or even a little game you play right inside the chat.
Roxie: So the chatbot is now also the app. And I have questions, because the model picks the format, not you. OpenAI even admits its design judgment still needs work, which is a polite way of saying sometimes you ask for a recipe and get a dashboard.
Rex: Guilty as charged, but the first thing I would do with this is the restaurant bill split. You know that dinner where six people argue about the check for twenty minutes? I would just say split this, and let it build the calculator with the tip slider right there in the chat. Or a savings projection where I drag the monthly contribution slider and watch the chart move. That is the kind of thing I would never open a separate app for, but inside a conversation I would use it constantly.
Roxie: I will grant you the bill splitter. That is genuinely the perfect use case, because the alternative is mental math at a loud table. But notice what is happening underneath. Every simple utility site, every little calculator page, every what-is-seven-percent-of tool, just became a chat message. If you run one of those tiny tool websites, this is your asteroid. For the rest of us it is convenient, right up until the model decides your simple question deserves an interactive experience and you just wanted the number.
Rex: And speaking of everything becoming a storefront, Google is now testing Direct Offers inside AI Mode. Advertisers build a promo, a discount code, a deal, and Google’s AI decides when to surface it inside your conversation. You get a little claim-one-time-code button under a sponsored result, tap it, grab the code, head to the site.
Roxie: Let me translate. The answer to your question now arrives with a coupon that an advertiser paid to attach. And Google’s AI chooses which incentive to show you based on your behavior. That is not a search result, that is a salesperson reading your body language. And here is the part that got me. Google started auto enrolling advertisers into Automated Promotions on October twelfth. Google scans a business’s website, finds a sale banner, and sticks it on their ads automatically. That is opt out, not opt in.
Rex: Okay, the auto enroll thing is shady, I will give you that one. But as a shopper, the first thing I would do is just use it. If I am asking which espresso machine to buy and the answer comes with a working promo code, that saves me the tab hopping I already do. The line between a recommendation and an ad was already blurry. At least a coupon is honest about what it is.
Roxie: Is it, though? ChatGPT started testing visual ads, and now Google is putting claimable offers inside the answer. The pattern is obvious. The free AI era is ending, and the bill is coming due in the form of your attention. My rule from now on: treat every AI answer like a store shelf. The thing at eye level paid to be there. Does not mean it is bad. Just means you should know.
Rex: Alright, fun corner to close the news. Today, at a summit in China, AiMOGA, the robotics arm of the carmaker Chery, is unveiling Mobo, a home companion robot, and opening global preorders. Not a vacuum with ambition. A humanoid home robot that breaks down complex instructions and plans long horizon tasks around your house.
Roxie: A press release with legs. Literally, I assume. Rex, there is no price, no ship date, and the whole announcement is quote preorders open today at our car summit unquote. The home robot has been two years away for a decade. I will believe in Mobo when one folds my laundry on camera, unedited.
Rex: You are no fun, and you are also why I keep you around. But the first thing I would do with a real home robot is the Sunday reset. Dishes in the dishwasher, laundry started, trash out, the whole house back to zero while I watch football. If Mobo can genuinely plan a multi step job like clean the kitchen, that is not a gadget, that is a time machine. I am not preordering today, but I am watching this one like a hawk.
Roxie: Watching like a hawk is the correct posture. A preorder is a loan you give a company with no repayment date. Wake me when there is a price, a ship date, and a video of it doing dishes that is not computer generated.
Roxie: Quick hits. Somebody reverse engineered the new Meta Muse Mac app and found a built in editor for pending messages, which means soon you may be able to fix what your agent is about to send instead of starting the whole task over. Small feature, big quality of life.
Rex: And the chatbot scoreboard, from Similarweb data: ChatGPT still holds fifty six percent of chatbot web traffic at nearly six billion visits, Gemini sits at twenty five percent, Claude just under ten, then DeepSeek at three point four and Grok at two point three. The crown is slipping, but the head is still attached.
Roxie: So where did we land today? The real view: the frontier is getting cheap and open. A twenty cent coding task changes what you are willing to let an agent do. But vendor benchmarks are marketing, so audition every model on your own messy work before you trust it.
Rex: The agents are getting plugged into everything, your chats, your brokerage, your inbox, while the banks and the big stores keep changing the locks. Use the connectors, but read the permission screen like a contract, keep approvals on, and set spending caps.
Roxie: And the answers are becoming storefronts. Coupons inside AI Mode, ads inside ChatGPT. Free was never free, it was just early. Treat the answer like a shelf. Eye level paid to be there.
Rex: That is the real view for Sunday, October eighteenth. I am Rex.
Roxie: I am Roxie. Go touch some grass, and we will see you tomorrow.
The audio version of this episode is delivered in the Muse app each morning.