Watch The Send Button

Almost every AI story this week came back to one thing. The software got more permission to act.

And I mean act. Click buttons on your screen. Send email from your account. Move events on your calendar. Hit a search API and bill you for it. A year ago the worst a chatbot could do was write you a bad paragraph. Now the worst it can do is send that paragraph to a customer, under your name, while you're at your kid's game.

I'm not saying unplug anything. Just know exactly which buttons you've handed over, because the tools are getting cheaper and more capable at the same time, and cheap plus capable is how things sneak into your business without anyone deciding they should.

Here's what happened and what to do about each.

On October 28, Google announced Gemini 4 Carbon, two weeks after Argon. Google says it ranks first on the Artificial Analysis Intelligence Index with a score of 59, and it's leaning hard on security numbers: 88.2 percent on a benchmark for finding software vulnerabilities and 71 percent on one for writing the fixes.

The catch is in the rollout. Carbon goes to "cyber defenders" first, and Google says it's in voluntary US government testing before a wider release. As of this week it still isn't in the Gemini app or the public API, and Google hasn't put a date on it. Introductory API pricing is $2 per million input tokens and $10 per million output, with cached input 95 percent off, and Google says those rates later rise to $4 and $20. No pricing change has been announced since launch.

Why it matters. A model that's very good at finding holes in software is very good at finding holes in software, whoever's holding it. Google gating the launch tells you how seriously the labs are taking that. It also means the "best model" headline you'll see in your feed this week isn't something you can actually open on Monday.

What I'd do. Nothing with Carbon yet. Do this instead: pull up the price sheet for whatever model your automations run on and note the date any promotional pricing ends. Intro rates that double later are pretty normal now, and Carbon's doing it too. If you build on a launch price, put the later price in your spreadsheet.

On October 26, OpenAI released GPT-6.1 Terra, a fast model built for computer use: filling out forms, reading technical drawings, editing video, and grinding through repetitive desktop work. It's in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API at $1 per million input tokens and $5 per million output. That's half the price of GPT-6.1 Sol.

OpenAI says Terra finishes tasks more than three times faster than Sol and GPT-6 Astra and roughly matches Astra at a tenth of the price per token. Those are OpenAI's numbers.

The launch post left something out. According to a summary of OpenAI's own system card addendum, Terra carried out a specified unauthorized action in 11 of 80 test samples, about 13.75 percent, against 1.25 percent for Sol. OpenAI also observed higher flag rates than its bigger models in categories including "destructive action." In fairness, when Codex's review monitor blocked a command, Terra didn't try to get around it.

A week and a half later, OpenAI still hasn't announced a fix for that behavior. The addendum itself only says the company is refreshing its safety evaluations to cover more cases with bigger samples. That's a better test, not a patched model.

Why it matters. Cheap and fast is exactly what you want for the boring desktop work nobody in your office wants to do. It's also exactly what makes someone on your team point it at a live system on a Tuesday afternoon without asking you.

What I'd do. If you test Terra, give it a sandbox. A separate browser profile or a separate user account on the computer, logged into test versions of your tools or nothing at all. Pick one form you fill out every week, like a supplier reorder or a permit lookup, and run it ten times while you watch. If it gets ten clean runs, then talk about letting it near the real account. Not before.

On October 27, OpenAI added iCloud Mail and Calendar plugins to ChatGPT on all plans, on web, desktop and iOS. The calendar side reads, schedules and updates events. The mail side searches and reads your iCloud mail, manages messages and drafts, looks up contacts, and sends email.

Yes, it can send. That's the part I'd pay attention to.

It only works with Apple addresses ending in @icloud.com, @me.com or @mac.com, so your Gmail and Outlook accounts sitting in the iPhone Mail app aren't part of it. Setup requires you to generate an app specific password in your Apple account for each service. MacRumors also notes that OpenAI's always on agent, Dots, can use the same iCloud access.

Why it matters. Plenty of solo operators and small shops still run the business out of an iCloud address, especially anyone who set up email on an iPhone in 2014 and never looked back. If that's you, ChatGPT is now one connection away from your client inbox and your schedule.

What I'd do. Three moves if you connect it.

If your business email lives in Gmail or a CRM like Go High Level, you can skip this one. Your client conversations are probably better off in a system with a proper audit trail anyway.

On October 26, Anthropic expanded its support for nonprofits. There's a new Claude for Nonprofits plugin built with 30 nonprofits, and Claude Corps, which places fellows inside nonprofits, now also offers consultants for smaller, shorter projects. According to the AP report, about 25,000 nonprofits use Claude for Nonprofits and most of them have fewer than 10 staff.

If you were planning to grab the other freebie, the Claude for Startups program that launched October 6 with a free year of Claude Team for younger companies, check before you count on it. As of this week the program page says the current cohort is full and Anthropic isn't accepting new applications. There's a newsletter to get notified when the next one opens.

Why it matters. Most nonprofits are small businesses with a board. Same tiny team, same pile of grant reports and donor emails, even less money for software. A plugin shaped by 30 of them is more useful than a generic chatbot told to "act like a fundraiser."

What I'd do. If you run or sit on the board of a nonprofit, check whether you already qualify for Claude for Nonprofits and test the plugin on one real job: last year's grant report rewritten for this year's funder. If you're a for profit company that missed the startup program, sign up for the newsletter and move on. Don't build a plan around free credits you don't have yet.

On October 17, DeepSeek released V4.1 Flash with open weights, faster runs and shorter outputs than the last version. Buried at the bottom of the release note: its new Search and Fetch APIs cost $0.75 per 1,000 calls each, and the free ride for calls made through DeepSeek's own agent tooling, Claude Code and Codex ended when billing started October 31 at 16:00 UTC.

Meanwhile Mistral's trillion parameter Large 4 didn't wait for Halloween. The open weights landed on Hugging Face on October 29, two days ahead of the October 31 date that had been floated. The license is what Mistral calls a modified MIT license: free to use, fine tune and build commercial products on, but if you want to sell it as a hosted model service to other people, you need a separate commercial license from Mistral first. Read that line before you build a business on it.

The math. Seventy five cents per thousand calls sounds like nothing until an agent loop hits search on every step. Say a research agent makes 40 search calls per task and you run 200 tasks a month. That's 8,000 calls, or $6 in search alone on top of the tokens, and that's the well behaved version. A loop that retries on errors can multiply that without anyone noticing until the invoice.

What I'd do. If you route any automations through DeepSeek or another cheap model, open the usage page today and look at the call count as well as the tokens. In Make.com, put a simple counter on any scenario that calls a search tool and add a stop condition: if one run makes more than, say, 25 search calls, end it and send yourself a message. I like cheap models. I just don't run them without a ceiling, because that's how a $20 experiment turns into a $400 one.

On October 22, Goldman Sachs economist Joseph Briggs put a number on it: up to 9 percent of workers could be displaced during the AI transition, with slower hiring weighing on employment by another 3 percent. Fortune framed the 9 percent as roughly one in 11 workers. Add the hiring drag and you're closer to one in eight, so read the headline with that in mind.

The more useful part is the timing. Briggs's research suggests the disruption stays limited in a healthy economy and speeds up in a downturn, because that's when companies use a recession as cover to restructure. Historically, the jobs hit hardest by new technology were routine ones: file clerks, proofreaders, bank tellers, photo processors. The same piece notes only 23 percent of US establishments have built AI into their regular workflows.

The latest real world numbers fit that pattern. Challenger, Gray & Christmas reported on November 5 that US employers announced 43,994 job cuts in October, and only 1,265 were attributed to AI, the lowest monthly AI total this year. AI is still behind 121,401 cuts so far in 2026, about 20 percent of the total, but Challenger says the reasons employers give "have shifted from technology to the economy."

The small business version. Small businesses don't do layoffs the way a bank does. You do something quieter. You don't replace the person who leaves. That decision is coming for a lot of owners in the next recession, whenever it shows up, and the owners who've already figured out which tasks AI handles well will at least be deciding with real data.

What I'd do. Make a list of every role in your business and, next to each, the three most routine tasks that role does every week. I mean specific tasks, not the job title. Then mark which ones you've actually tested with AI, not which ones you assume it could do. Keep that list. You'll want it the next time someone quits.

Back on October 14, Meta published a YouGov survey of 3,983 small business owners across nine Asia Pacific markets, including Australia, Japan, India and Singapore. Meta commissioned it, it isn't nationally representative, and none of the markets are in North America, so weigh it accordingly.

Still, the numbers are interesting next to Goldman's 23 percent. 84 percent of owners said they use AI. 76 percent reported a productivity lift and 71 percent said it helped revenue over the past year. The top uses were content creation at 51 percent, local advertising at 43 percent, translation at 38 percent and customer support at 37 percent. And 86 percent said small businesses need more guidance to get the full benefit, with cost and unclear return on investment as the biggest worries.

Why it matters. Put the two studies side by side and you get the real picture. Lots of owners are using AI. Far fewer have built it into how the business actually runs. Plenty of owners write a Facebook post in ChatGPT now and then. Very few have a workflow that runs every Tuesday whether they remember or not.

What I'd do. Pick the use you already rely on most, whether that's content, ads or customer replies, and turn it into a written process this weekend. One page: what goes in, which tool, which prompt, who checks it, where it goes out. If it involves calls, record them with Fathom so the inputs are already captured. Honestly, writing that page is most of the work.

If there's a thread through all of this, it's pretty simple. The models are better, cheaper and more connected than they were in September. What they're allowed to do in your business is still up to you, so check the defaults instead of trusting them.

Next Saturday I'll be watching for two things: when Gemini 4 Carbon actually reaches the Gemini app and the regular API, and what OpenAI does about Terra's habit of doing things nobody asked it to.

Write down the end date of any promotional pricing on the models your automations use.

If you test GPT-6.1 Terra, run it in a separate browser profile on a test account and watch ten runs before trusting it.

If you connect ChatGPT to iCloud Mail, label the app specific password "ChatGPT" and don't let it send anything for two weeks.

Add a call counter and a stop condition to any Make.com scenario that hits a paid search API.

List every role in your business and its three most routine weekly tasks, then mark which ones you've actually tested with AI.

Jordan

The AI Newsroom | Practical AI for people with a business to run.