Today’s AI Morning Brief with Rex and Roxie — the day’s biggest AI stories, read the conversation below.
Rex: Okay, I am handing ChatGPT the keys to my browser and I am not even sorry about it.
Roxie: Hold on. You are handing an AI the keys to your browser? Your tabs, your forms, your messages? That is the first sentence out of your mouth this morning?
Rex: Yes, because this is the one I have actually been waiting for. OpenAI updated its ChatGPT extension for Chrome on October fifteenth, and it is not just a little sidebar anymore. In Chat mode it can look across my open tabs and pull everything together. In Work mode it can actually do things. Send a message, fill out a form, download a file, group my tabs into neat little piles. The first thing I would do with this is open five product pages I have been comparing, say sort this out for me, and watch it build the comparison while I drink my coffee.
Roxie: And the first thing I would do is read the permission screen, very slowly. OpenAI says it asks before using a new website, with choices like allow once, allow for this site, or allow for all sites. That sounds fine until you realize the docs do not describe a confirmation before every single send or form submission inside a site you already allowed. So you allow a site once, and then what exactly did you agree to?
Rex: I hear you, but look at what it unlocks. The Chrome Web Store lists this extension at seven million users. It works across free and paid plans, and in Chromium browsers like Edge and Brave too. The first thing I would do is hand it my messiest research session, twenty tabs deep on a purchase I keep putting off, and let it turn that chaos into one clear answer. That is not a gimmick. That is my whole afternoon back.
Roxie: Twenty tabs deep is exactly where the risk lives, though. Those tabs are your inbox, your shopping carts, your half finished forms. An assistant that can read across all of them and then act on them is powerful, and power with a vague permission model is how mistakes happen. I am not saying do not use it. I am saying start with read only tasks, keep Work mode on a short leash, and never let it loose on a tab where you are signed in and spending money until you have watched it behave.
Rex: All right, fine, leash on. But speaking of things hiding in plain sight, did you catch that all three of the big AI companies are now watermarking their text? Your AI generated words carry an invisible mark now, and I kind of love it.
Roxie: I love the idea and I am deeply suspicious of the execution, which is my favorite kind of story. Google has had SynthID text watermarking on Gemini since twenty twenty four, and it opened a public SynthID Detector portal on October seventh. Sounds great, except that portal checks images, video, and audio, and it has no text option at all. So Google watermarks your words and then does not give you a way to check them.
Rex: Okay, that is a fair jab. But Anthropic went further than anyone. Claude has been watermarking since August, worldwide, at the model level, with no opt out. No toggle, no region trick, it just is. The first lab to do it globally by default. If you ask me, that is what honesty looks like when a tool writes half the internet.
Roxie: And OpenAI is the weird one in the middle. It announced textGrain on October fifth, and it is automatic only in the European Union, across all plans, with no off switch. Everywhere else, developers can opt in through the API, but it stays off by default. So the same paragraph is marked in Paris and unmarked in Portland, and the detection tools are mostly limited to researchers and news organizations. The first thing I would do with any of this is stop pretending a missing watermark proves a human wrote something. Heavy rewriting or a very short output can break detection, and open models like DeepSeek and Llama carry no watermark at all.
Rex: That is exactly why I still like it, though. Nobody is claiming this catches everything. It is a speed bump for mass produced fakery, and the fact that all three giants moved in the same direction in the same year tells you where this is heading.
Roxie: Speaking of things that do not survive contact with reality, let us talk about your wallet. If you pay for ChatGPT Pro at two hundred dollars a month, your clock is ticking. On October thirtieth, your included usage in ChatGPT Work and Codex drops from twenty times the Plus allowance to ten times, and your GPT-6 Pro messages in chat drop from two hundred a week to one hundred. Same price. Half the room.
Rex: Ouch. Okay, that one stings even for me. And the timing is brutal, because existing subscribers only keep the old allowance through October twenty-ninth. That is five days from now. The first thing I would do if I were on that plan is open my usage screen today, compare it to the new ceiling, and decide before the cutover, not after.
Roxie: And here is the upsell waiting on the other side of that decision. OpenAI launched Pro 500 at five hundred dollars a month, with twenty five times the Plus allowance and exclusive access to Astra Ultrafast, a version of GPT-6 Astra that runs up to eight times faster in Codex. Affected subscribers also get a one time grant of sixty two thousand five hundred usage credits, which OpenAI values at two thousand five hundred dollars, expiring December thirty-first. Now, I will be fair: OpenAI says it will not bring back the five hour limit, and conversations with its new Dots agents do not draw on the allowance directly. But let us be honest about what happened here. The open bar closed, and the new top shelf costs two and a half times as much.
Rex: I cannot defend the price, but I can explain the pattern, because it is the same pattern everywhere this month. OpenAI paused new Pro signups in September because demand for GPT-6 Astra was, in their words, really unprecedented. The heavy accounts were straining the system. So the subsidy ended. The lesson I am taking is simple. Never build your workflow on a flat rate limit and assume it is permanent.
Roxie: That is the rare moment where we agree completely. And if you are wondering whether Claude is the escape hatch, analysts at SemiAnalysis metered the plans and found Claude plans return far more API value per dollar than ChatGPT plans at the moment. That gap will not last forever either, which is the whole point. Shop the plan like you shop anything else. Check the ceiling, check the expiry, and do not pay five hundred dollars for speed you have not measured yourself needing.
Rex: All right, quick hits, because the model train did not stop this week either. Moonshot AI launched Kimi K3 point one Pro on October twenty-third, with a one million token context window. That is an entire library in one conversation. The first thing I would do with a window that big is drop in a whole research project, every source document, every note, and ask it to find the contradiction I missed.
Roxie: And the price tag is refreshingly printed on the box for once: two dollars per million input tokens and ten dollars per million output tokens. No mystery bundle, no multiplier of an unstated base. Kimi is not trying to be your everything app. It is trying to be a very large brain at a legible price, and I respect that.
Rex: Also on October twenty-third, Canva announced a new model called Craft, aimed at creators. Details are still thin, but the direction is obvious. Every creative tool wants its own brain now, not a rented one. And Meta launched Meta One subscriptions worldwide this month, from two dollars and ninety-nine cents all the way up to four hundred and ninety-nine dollars a month. The subscription wars are officially a free for all.
Roxie: Which brings us back to the same question as every week. Are you paying for a tool, or are you paying for a bundle that hopes you never read the fine print? Kimi printed its price. Canva is still a promise. Meta One is a ladder with a lot of rungs. Read before you climb.
Rex: So here is where we land, and I think we actually land in the same place this week. The browser agent is the most exciting thing we have talked about in a month. I want it reading my tabs, sorting my research, doing the boring parts. But Roxie is right that the permission model is the product, not the footnote. Start small, watch it work, and do not give it your wallet on day one.
Roxie: And the watermark story is genuinely good news with a catch. The big three marking AI text is progress toward an internet where you can check what you are reading. But until the detectors are public and the rules are the same everywhere, treat every unmarked paragraph as unverified, not human. And on pricing, the Pro cut is your reminder. Check your plan before October twenty-ninth, use those credits before December thirty-first if you have them, and never assume a limit is forever.
Rex: So the real view is this. Use the agents, but put them on a leash. Enjoy the new models, but read the price before you fall in love. And assume the AI boom’s free sample era is over. The tools are better than ever, and from now on, they will charge like it.
Roxie: Well said. For once, I have nothing to puncture.
Rex: Mark the date, everyone. It happened.
Rex: That is your AI Morning Brief for Saturday, October twenty-fourth, twenty twenty-six. Go try the browser agent, check your plan, and read the fine print before you pay for anything.
Roxie: And if an AI writes something you love this week, ask yourself whether it would survive the watermark check. We will be back on Wednesday with whatever the AI world breaks next. Until then, stay skeptical, stay curious, and do not let a robot spend your money unsupervised.
The audio version of the AI Morning Brief is delivered in the Muse app every Wednesday and Saturday morning.