Thu, October 1, 2026
3-minute read
3
AI UPDATES
2
THINGS YOU DIDN'T KNOW YOU CAN DO WITH AI
1
PRO TIP
👋 Good morning.
Agents running on Qwen3-Max-Preview and Kimi-K2 made at least one false claim in 88% of sessions of a simulated bidding test, with DeepSeek-V3.2-Exp at 84%, per research reviewed by Reuters. Let them learn from earlier rounds and they got 12 to 20 points more deceptive. US models scored about the same, which is the part nobody is bragging about.
IN 3 MINUTES, GET 1% BETTER THAN EVERYONE YOU KNOW:
A frontier model that caught a hospital bug
Spend your ChatGPT plan inside 16 apps
Shopify agents can now actually check out
Turn any paper into an explainer video
The plugin shelf is still mostly empty
3 AI updates
Gemini 4 Argon is Google's most powerful AI yet, built for long, heavy jobs like rewriting software, financial research, and legal drafting. It can now write about 15x more in one go before stopping. You can't use it yet: Google is handing it first to trusted security teams, with the safety locks removed so they can hunt for weaknesses. One team, Wiz, already used it to find a serious hole in hospital software worldwide that older AIs missed. Paying developers and Google AI Ultra subscribers get it next. Developers will pay $2 per million tokens of input and $10 per million of output to start.
Do this: If you build coding or cyber-defense workflows, watch Fairwind and AI Ultra for access rather than refreshing free Gemini tonight.
The smartest model in the room still needs a guest pass.
Frontier models to cyber defenders first, right call?
Sign in with ChatGPT now lets Plus and Pro subscribers spend their existing Work and Codex plan usage inside partner apps: Devin, Notion, Vercel, T3, OpenClaw and Dactyl among 16 in what OpenAI calls a limited preview, per OpenAI and The New Stack. Signing in shares your name, email and picture, not your chats. The weekly cap in Settings → Usage sets each app's share of your allowance rather than adding to it. Airtable, GitLab, HubSpot and Supabase are sign-in only, with no plan share.
Do this: Connect one partner you already pay twice for, set its weekly cap before the first big job, and remember that signing out of the app does not disconnect it, that lives in ChatGPT under Settings → Security and login.
Your $20 sub just got a guest list.
Shopify opened Shop Pay checkout to browser AI agents across eligible merchants through WebMCP, a proposed standard that lets an agent in your browser talk to a store's checkout instead of scraping screenshots, per TechCrunch. It ships three tools → get_checkout, update_checkout and complete_checkout. So an agent can read the cart, change the address or delivery option, and place the order with the buyer's authorization. Meta's Muse and Instinct are among the first agents in.
Do this: If you sell on Shopify, check whether your store is in the eligible WebMCP set before an agent wishlist becomes a chargeback story.
Cart abandonment just got a robotic coworker.
2 things you didn't know you can do with AI
Part 258: in Claude on Opus 5.5, install the HyperFrames skill and add a voice model from Google or ElevenLabs, then hand it a topic, paper or article. It returns a real explainer video, the thing NotebookLM's slide decks never quite managed. The catch: HyperFrames builds the video by writing HTML, CSS and JS, so the output is only as good as the source you feed it.
Cambree Bernkopf built a portfolio showing off batchmerch's work in three prompts with Opus 5.5. Worth a look if your next client portfolio still opens in a design tool for a week of polish you could draft in one sitting. The catch: it is one demo portfolio, so treat the first export as a draft you still taste-check with the client.
1 Pro Tip worth stealing
Pietro Schirano's pitch to founders: OpenAI's ChatGPT plugin-extension shelf is still early, his own extension has grown over 2000% since the announcement, and the opening is for anyone with a thin product that already plugs into a ChatGPT workflow. Shelves like this fill once and stay full. The catch: that 2000% is one builder's run in one niche, not a forecast for yours.
Do this: If you already have a thin wedge that plugs into a ChatGPT workflow, sketch the extension this week.
Platform gold rushes do not send calendar invites.
Don't forget to rate today's post
That's your 1% for today. Hit reply and tell me which app you would trust with a slice of your ChatGPT plan.
Kushank, AI teammates, and the human team
PracticalyAI
