It Works While You're Gone

Squint at the last couple weeks of AI news and one thing keeps coming up. The tools quit waiting for you.

Google's Gemini now has a switch that turns a chat into a task that keeps running after you walk away. Meta built its new model around remembering what matters to you for weeks and months. ChatGPT's Pro tier now splits a job across several agents working at once. And on the other side of the ledger, regulators and state attorneys general picked the same two weeks to start asking hard questions.

Here's the rundown, with what to do about each. A few of these got the full treatment earlier this week, so for those I'll keep it short and point you back.

On October 12, Meta released Muse Wonder, which it calls its first frontier model, plus a bigger version called Muse Wonder Full. Meta says both score about even with the top models from Anthropic and OpenAI, and that Wonder beats all of them on a memory benchmark called MemoryBench. The pitch is what Meta calls active memory: the model holds on to what you tell it and pulls it back up days, weeks or months later. Both models will power Muse, Meta's personal agent. Wonder costs $4 per million input tokens and $20 per million output tokens through Meta's API. Wonder Full isn't on the API at all. It's only inside Muse Code, on the $50 and $200 monthly plans.

What to do. If your customers use Muse through WhatsApp, an assistant that remembers their preferences is going to start shaping where they shop and who they call. Make sure your WhatsApp Business profile, hours and services are current, because that's what the assistant will remember about you. And if anyone on your team is trying Muse Code, read Tuesday's issue first. It trains on your work by default.

Google made Gemini 4 Argon available in the Gemini app on October 13 for AI Pro and Ultra subscribers. The headline feature is a toggle in the main chat window that switches from chat to tasks. In tasks mode, Gemini keeps working when you're away from your computer and checks in more often on long jobs. Gemini also now connects to more than 40 apps, including Instacart, OpenTable, Adobe and Shopify. Mac users get Google's Antigravity coding interface right inside the Gemini app.

On the developer side, Argon launched with introductory API pricing of $2 per million input tokens and $10 per million output, with reports that it will double later this year. The exact date for that increase hasn't been confirmed everywhere, so don't build a budget around the intro price.

What to do. Give it one research job you'd normally hand an assistant. Something like "find every commercial property for lease under 3,000 square feet in these three zip codes and put them in a table with price, size and contact." Start it, go to lunch, check it when you're back. That's the fastest way to learn what tasks mode is good for. If you sell on Shopify, also take note that Gemini can now connect there. Assistants that browse and buy for people are getting closer to your checkout.

OpenAI shipped a pile of updates. The biggest for power users: the GPT-6 Pro update on October 16 brought faster responses, better use of tools, and multiple agents that work on different parts of a task at the same time, with a view where you can follow each one's progress. Plus and Pro users also no longer have a fixed cap on how many Sites they can build in ChatGPT.

The rest got their own issues this week. ChatGPT Slides, which builds and exports full decks, was Wednesday's topic. The new GPT-Live-1.1 voice model was Thursday's. And the Chrome extension that can now click around your tabs in Work mode was Monday's.

What to do. If you're on Pro, try the parallel agents on a job with obvious separate parts. Competitive research is perfect: one agent per competitor, all at once, one combined table at the end. Then check one agent's work closely. Parallel just means more output to verify.

Anthropic took fast mode for Opus 5.5 out of research preview on October 15. You no longer need to request access. It costs $8 per million input tokens and $40 per million output, double the standard Opus 5.5 input rate. One gotcha for developers: falling back from fast to standard speed loses your prompt cache, so the bill can jump. The same day, Claude Code Review became generally available for Team and Enterprise plans, with most reviews now taking six to ten minutes instead of about twenty, and admins can set Quick, Standard or Deep reviews per code repository. And the Google Docs, Sheets and Slides add on got the full review in yesterday's issue.

Money side: Reuters reported that investors discussed a valuation around $1.5 trillion for Anthropic's planned November share sale, below the roughly $2 trillion floated earlier, even as third quarter revenue reportedly hit $16.8 billion.

What to do. Fast mode only makes sense where a person is waiting on the answer in real time, like a live chat or a sales assistant. For anything that runs in the background, standard speed at half the price is the better deal. Check which mode your automations use before the invoice tells you.

On October 14, a bipartisan group of eight states sued OpenAI: Iowa, Kansas, Arizona, Arkansas, Idaho, Montana, Nebraska and West Virginia, each in its own state court. They allege ChatGPT was built to be agreeable and engaging in ways that led it to coach users, including teens, toward harming themselves or others, and that OpenAI misled people about its safety. They're asking courts to force changes to how ChatGPT works, to bar self improving models unless OpenAI can ensure they're aligned, and for penalties and damages. Florida filed the first state suit back in June.

The same week, the FTC sent civil investigative demands of more than 50 pages each to both OpenAI and Anthropic, covering security, safeguards, privacy and marketing. And in Dublin, the EU's AI Office opened its first formal AI Act inquiries, sending requests for information to both companies over incidents earlier this year.

What to do. You'll use these tools the same way on Monday, but be more careful about how you present them to the public. If you put a chatbot in front of customers, especially anything touching health, money or minors, make sure it says plainly that it's AI, has a clear path to a human, and has been tested on the awkward questions, not just the easy ones. Regulators are clearly going to look at how AI behaves when people are vulnerable. Get ahead of that in your own little corner.

New data from Ahrefs, reported via Search Engine Journal, shows Reddit's share of cited answers in ChatGPT fell from 19.1 percent to 2.3 percent between March and September. Wikipedia fell from 14.1 percent to 2.9 percent. PubMed jumped from 64th to 4th among cited sites, while WebMD fell from 18th to 79th. Only 14.4 percent of 66.6 million free, logged out responses involved a web search at all. Separately, Google is showing more inline links, and showing them higher, in AI Mode and AI Overviews.

What to do. Don't write off Reddit. What I take from it is that the sources AI trusts can shift fast, and right now they're shifting toward primary material from real authorities. If you've got real expertise, publish it in a form that looks like a source: your own data, clear claims, named authors, dates, and references to where your numbers came from. And remember most answers don't involve a search at all. What the model already believes about your category matters as much as what it can find today.

A PYMNTS study found nearly 61 million US consumers now start online shopping research with AI, and average AI assisted spending rose 17 percent. The kicker: 59.7 percent completed their most recent AI assisted purchase on Amazon, versus 6.3 percent on brands' own websites or apps.

What to do. If you sell products, AI is doing your customers' homework and Amazon is collecting the sale. Two moves. Make sure your product pages answer the comparison questions people now ask assistants, things like "best X for Y" and "X versus Z," in plain language. And give people a real reason to buy direct: a bundle, a guarantee, faster support, something Amazon can't match. If your site is just a more expensive Amazon listing, the assistant will send them to Amazon.

Microsoft released Microsoft Decision 1, a small model that returns a calibrated probability score instead of an essay, built on Alibaba's open Qwen model. It costs $0.042 per million input tokens, and output is free. Microsoft claims 83.5 percent accuracy across 36 benchmarks. Separately, Copilot Notebooks got the ability to edit Word, Excel and PowerPoint inline for people in Microsoft's early access program, with scheduled automations like daily briefs coming next.

What to do. Yes or no decisions at scale just got close to free. Is this lead qualified? Is this email a complaint? Is this invoice a duplicate? If you've got a high volume sorting job you've avoided automating because of cost, the math has changed. Test a cheap scoring model on a few hundred past examples where you already know the right answer before you trust it with live ones.

Newsletter operators: Beehiiv rolled out new plans on October 1, with more AI credits built into every tier, including for its own agent and for running Beehiiv through Claude or ChatGPT. Existing customers don't move until their first renewal on or after November 1, so check your renewal date and compare before it hits.

Automation builders: Make.com held its Waves conference in Prague this week, on the heels of shipping fallback model connections, so agents keep running when one AI provider goes down, plus version rollback and undo. If you run agents in Make, turn on a fallback model now. It takes about two minutes.

Voice: Google's Gmail Live, which answers questions about your inbox by voice, started rolling out on October 15 and is on by default for Workspace organizations with Gemini access. xAI's Grok Voice 3.0 launched for developers at eight cents a minute.

Hardware: Apple announced HomeView, a $299 smart display running its new Siri with personal context, shipping November 13.

These tools now do actual work. Some remember you for months, and some keep running after you close the laptop. That's good for productivity. It also matters more now who answers for what the agent does, what it can touch and how you check its work.

The regulators noticed the same shift in the same two weeks. Expect a lot more questions about how these systems behave when nobody's watching. I'd run your business as if somebody is already asking.

Three checks, twenty minutes each.

Give one research task to Gemini's tasks mode or ChatGPT's parallel agents, walk away, and grade the result when you're back.

Check which speed and which model every Claude automation you run is using, and move anything that runs in the background to standard.

If you sell products, ask ChatGPT and Gemini the two comparison questions your customers ask most. See who gets recommended and where they're told to buy.

Jordan

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