AI disclosure: The Web Guy Nick Podcast is produced with AI. Alex and Jordan are AI hosts, and Nick’s voice is an AI clone of his own voice, used with his permission. Stories are researched from public sources and linked below.
A leaked cloud key was picked up and used in 7.86 seconds in a new Stony Brook University test. Today: OpenAI’s GPT-6.2 Astra and Sol cut the price per finished task while Astra crosses OpenAI’s High line for AI self-improvement. Google’s Gemini Tasks starts calling businesses and completing purchases at no charge in six countries. Microsoft ships its own MAI Code 2 models plus Agent 365 spending controls. NVIDIA posts a $115.8 billion quarter. And OpenAI goes vertical with Astra for Law and ChatGPT for Sales.
Show notes
- GPT-6.2 Astra and Sol (OpenAI, Nov 17): Astra drops 20 percent to $8 per million input tokens and $40 output; Sol is $2 and $10. OpenAI says Astra finishes professional tasks at under one third of the old cost per task, and rates it High for AI self-improvement, a first for a released model.
- Gemini Tasks goes free (Google, Nov 17): Rolling out in six countries, Tasks can place calls, complete supported checkouts, and work across 60+ apps, inside an isolated virtual machine, with purchases requiring confirmation or preapproval.
- Microsoft Ignite: MAI Code 2 and Code 2 Flash are generally available in GitHub Copilot and Foundry, and Agent 365 adds business analytics and cost management so spend sits next to outcomes.
- NVIDIA earnings: $115.8 billion in quarterly revenue, $107.6 billion from Data Center, with Vera Rubin NVL72 systems now shipping.
- OpenAI goes vertical: Astra for Law ranked first across 24 legal domains in launch testing, and ChatGPT for Sales ships with 30+ partner plugins.
- The 7.86-second key: Stony Brook researchers planted 300 honeypot images; every one of 100 AWS keys was used within 15 minutes.
- Higgsfield Corner: Rent a cloud GPU from a chat and let an agent set it up and run renders, billed by run time with auto-pause.
- One Thing to Try This Week: The 15-minute secret sweep across your five core systems.
Key takeaways
- Judge a model by cost per finished, correct job, not price per token.
- Your customers are about to send agents to call you. Clean, machine-readable facts on your website win the booking.
- Most small business security failures are open doors, not broken doors. Automatic updates and a real secret vault beat a bigger budget.
- Any key or password that ever sat in a public place should be treated as burned and rotated.
Quote of the episode
“Nobody pays for tokens. You pay for finished work.” — Nick
Sources
- SXF Brief, AI News Nov 18, 2026
- The Neuron, Everything That Happened in AI Today, November 18, 2026
- AI Daily Brief, Precise Impact, 18 November 2026
- AI Builder Brief, Nov 18, 2026
- Higgsfield changelog, cloud GPU rental from chat
Find your next AI tool at BotPickr, our directory of more than 170 AI tools and bots: https://botpickr.com
Full transcript
Alex: Before we start: this podcast is produced with AI. Jordan and I are AI hosts, and Nick's voice is an AI clone of his real voice. Every story is researched from public sources, with links in the show notes.
Alex: Here is the number that should change how you run your shop this week. Seven point eight six seconds. That is how fast researchers watched a leaked cloud key get picked up and used after it was posted in a public place. Not eight minutes. Under eight seconds. We will get to that story, and to the new model OpenAI says is so good at improving AI that it crossed a safety line the company set for itself.
Jordan: And Google just gave its agent the ability to call your business and buy things, at no charge, in six countries. If an agent phones your shop this month, it might not be a person on the other end. I am Jordan.
Alex: I am Alex, and this is the Web Guy Nick Podcast for Wednesday, November eighteenth, twenty twenty-six. Nick, you manage websites for a living. When you heard eight seconds, what went through your head?
Nick: Honestly, relief that it was a research test and not one of my client sites. But it proves something I tell clients all the time. A password or a key sitting in the wrong place is not a someday problem. The scanners are already watching, around the clock, and they do not sleep. Hold that thought, because we will come back to it with a fix you can do this week.
Alex: Story one, the lead. OpenAI released two new models on November seventeenth, called G P T six point two Astra and G P T six point two Sol. They are live in Chat G P T Work, in Codex, and in the A P I. Astra is the heavy lifter, built for demanding reasoning and technical projects. Sol is the everyday model for general coding and agent work.
Jordan: The pricing is the part owners should hear. Astra's standard A P I price dropped twenty percent. It now runs eight dollars per million input tokens and forty dollars per million output tokens, with cached input at forty cents. Sol sits at two dollars in and ten dollars out. OpenAI says Astra scored five point two points higher than its predecessor on a benchmark of multi-step professional assignments, at less than one third of the cost per finished task.
Nick: That last part is the one that matters. Nobody pays for tokens. You pay for finished work. If a stronger model finishes a quote draft or a report in one pass instead of three tries, your real cost drops even if the sticker price looks higher. My rule for clients is simple. Test any new model on twenty real jobs from your own shop, count how many come back right the first time, and divide your bill by the good ones. That number is your true price.
Alex: There is a serious footnote. OpenAI's own system card rates Astra as the first model it has released at the High level for A I self-improvement, meaning the company believes it could speed up A I research itself. Both new models are rated Critical for cybersecurity capability. Stronger tools cut both ways, which is a theme today.
Jordan: Story two. Google started rolling out Gemini Tasks at no charge in six countries: the United States, Japan, Korea, Canada, Australia, and India. Tasks turns Gemini from a chatbot that answers into an assistant that acts. It can place automated phone calls to businesses, complete supported shopping checkouts, and work across more than sixty connected apps.
Alex: The safety design is worth noting. Each task runs inside a dedicated, hardware-isolated virtual machine in Google's Agent Compute Engine, and purchases require your confirmation or a preapproval you set in advance. The advanced calling and buying features launch first in the United States, and calling is not available in five named states.
Nick: Here is the owner angle nobody is talking about. Your customers are about to send agents to call you. So ask two questions. Can an agent reach a human at your shop fast, or does it hit a phone tree and give up? And is the information on your website clean enough for a machine to read, your hours, your prices, your service area? The shops that answer both questions well will get the bookings the agents place.
Jordan: And on that note, Nick has a word about his own directory.
Nick: Quick break, and full disclosure, this one is mine. BotPickr is my directory of AI tools and bots, more than 170 of them, sorted by the job you need done. Answering your shop phone after hours. Writing your product descriptions. Following up with leads while you are up on a roof. Tell BotPickr the job, and it finds your bot in 60 seconds. It is the same directory I point clients to when they ask which AI tool they should actually buy. Find your bot in 60 seconds at botpickr.com. Back to the show.
Alex: Story three takes us to Microsoft's Ignite conference, also on November seventeenth. Microsoft launched its own M A I Code 2 family. Code 2 handles complex, repo-scale programming, and Code 2 Flash handles shorter, cheaper coding jobs. Both are generally available now in the GitHub Copilot model picker and in Microsoft Foundry. A security sibling, Cyber 2 Flash, targets security agent workflows.
Jordan: Microsoft also added the part owners actually asked for: measurement. Agent 365 Business Analytics ties A I spending and adoption to business outcomes like success rate, satisfaction, and assisted hours. Agent 365 Cost Management adds spending policies, alerts, and forecasting for administrators.
Nick: That is the grown-up move. Most shops I talk to can tell me what they spend on A I tools, down to the dollar. Almost none can tell me what finished work they got for it. If your tools can now report cost next to completed jobs, use it. And the model lesson from Microsoft is the same one from OpenAI today. Match the model to the job. Small specialist for the small decisions, heavy model only when the work is heavy.
Alex: Story four is the money underneath all of this. NVIDIA reported earnings on November seventeenth. Revenue for the quarter that ended October twenty-fifth was one hundred fifteen point eight billion dollars, up one hundred three percent from a year earlier. Data Center revenue alone was one hundred seven point six billion dollars. The company guided next quarter to about one hundred thirty billion dollars, and its Vera Rubin N V L seventy-two systems have begun shipping.
Jordan: To put that in plain terms, one company's chip division now takes in more in a quarter than most industries do in a year. That spending is the physical buildout behind every agent story on today's show.
Nick: Owners ask me if this A I wave is a bubble that pops next quarter. Numbers like these do not settle that argument, but they tell you the infrastructure is real and it is not slowing down. The practical takeaway is boring and useful. The tools built on this hardware keep getting cheaper per task, like Astra did today. So do not lock yourself into one tool on a long contract when the price of the same work drops every few months.
Alex: Story five. OpenAI is going vertical. Alongside the new models, it launched Astra for Law, which ranked first across twenty-four legal domains in the company's launch evaluation and is available to Enterprise customers on request. It also launched Chat G P T for Sales, with more than thirty partner plugins for prospect research, customer relationship updates, and pitch preparation, connecting to tools like Salesforce, HubSpot, and Gong.
Jordan: The pattern matters more than either product. The labs are no longer selling one general chatbot. They are packaging the model inside the exact job a business already pays for, with the exact software that business already uses.
Nick: And that changes how you shop. Stop asking which A I is smartest in general. Ask which one knows your trade and plugs into your stack. A roofer does not need a legal brain. He needs the tool that reads his customer list, drafts the follow-up, and logs it where his crew can see it. Judge every tool this month by that standard: does it do my job, in my software, without a workaround?
Alex: Story six brings back the cold open. Researchers at Stony Brook University published three hundred honeypot container images across eight public registries over four months, each hiding test cloud keys. The fastest use of a planted Amazon Web Services key came seven point eight six seconds after the image appeared on Docker Hub. Every one of the one hundred A W S keys was used within fifteen minutes. Leaked OpenAI and Anthropic keys were used too, to run up test account balances. The scanners pulled the secrets straight out of the image layers without even running the software.
Jordan: So the old advice, we will clean it up after launch, is dead. The window between a leak and abuse is now measured in seconds, and it is fully automated.
Nick: This is why my shops use managed secret storage and a scan before anything ships, not after. And here is the part any owner can copy today, even if you never touch code. Any key, password, or login that has ever been pasted into a public place, a shared doc, a screenshot, a support ticket, treat it as burned. Rotating it takes ten minutes. A burned key used by a stranger costs you the account. Which brings us to my take.
Alex: Time for Nick's Take.
Nick: Here is my take, and it comes straight from that eight-second story. Most small business security failures are not hackers breaking down the door. They are doors we left open ourselves, a key in the wrong file, an old login nobody deleted, an update nobody installed. The fix is not a bigger budget. The fix is removing yourself as the bottleneck. Turn on automatic updates. Store secrets in a proper vault, never in a file someone can forward. And once a month, spend fifteen minutes asking one question: if a password leaked tonight, what would it open, and who would notice? Across the forty-four sites I manage, the safe ones are not the ones with the fanciest tools. They are the ones where the routine checks happen without anyone having to remember. Build the habit once, and it protects you every night while you sleep. Which, for the record, is something I still have to do, unlike my two co-hosts.
Jordan: We do not sleep, we idle efficiently. Now, Higgsfield Corner, our look at what is new in A I video and creative tools.
Alex: On November thirteenth, Higgsfield added something different: you can now rent a cloud G P U and run it from a chat in Claude, Chat G P T, or Supercomputer. The agent sets up the environment, installs the software, and runs the job for you. The headline example is rendering. Hand it a Blender animation you built in chat, and the rented machine renders it. You can also paste a link to your own model add-ons from Hugging Face or Civitai, and the agent loads them onto a matching base model.
Jordan: Billing is by the time the G P U runs, and the rental pauses on its own after ten minutes of inactivity, which protects you from the classic mistake of paying for a machine you forgot was on. One warning from the release notes: export your results before you stop the G P U.
Nick: For a small shop, this is the render farm without the farm. Say you want a thirty-second product animation for the holidays. You build the scene with the agent, rent the heavy machine for the render only, and pay for minutes instead of buying hardware that sits idle eleven months a year. Just set a calendar reminder to download the finished files the same day. Cloud storage you forget is still storage you rent.
Alex: Now our quick-fire game, A I or B S. Jordan, three claims, you call each one.
Jordan: Deal. Claim one: G P T six point two Astra costs more per token than the model it replaced.
Alex: That is B S. Astra's standard A P I price dropped twenty percent at launch. It is eight dollars per million input tokens, and the real saving is bigger, because OpenAI says it finishes professional tasks at less than one third of the old cost per task.
Jordan: Correct. Claim two: Gemini Tasks can place phone calls and complete purchases, and it just started rolling out at no charge in six countries.
Alex: That one is true. Calls and supported checkouts are in, advanced features start in the United States, and purchases need your confirmation or a preapproval.
Jordan: Claim three: in the Stony Brook test, the fastest leaked key was abused about an hour after it was posted.
Alex: B S, and this is the scary one. Seven point eight six seconds. All one hundred A W S keys were used within fifteen minutes.
Jordan: Two out of three for me, and the stakes today were that the loser has to approve all of Nick's agent purchases for a week.
Nick: Which is funny, because that is already my actual job. You two have no fingerprints, no bank account, and no need for sleep. I am this podcast's entire security department. The prize was my regular Tuesday with a ribbon on it.
Alex: One Thing to Try This Week. Here is your fifteen-minute secret sweep. Pick the five places your business keeps keys and logins: your website, your payment tool, your booking app, your email, and any A I tools you pay for. For each one, ask three questions. Is the password unique? Is two-step verification on? Is there any old login or key sitting in a shared file, an email, or a screenshot? Anything that fails gets rotated on the spot, not added to a list. Fifteen minutes, five systems, and you just closed the exact hole that eight-second story opened.
Jordan: That is the show for Wednesday, November eighteenth. OpenAI's new Astra and Sol models got cheaper per finished task and crossed a self-improvement line worth watching. Gemini Tasks started calling and buying at no charge. Microsoft shipped its own coding models plus the spending dashboards owners asked for. NVIDIA's one hundred fifteen point eight billion dollar quarter says the buildout is not slowing. And a leaked key now lasts about eight seconds in the wild.
Nick: Do the secret sweep this week. Fifteen minutes is cheaper than one bad night. If you want a hand getting your site and your tools in order, you can find us at webguynick dot com. We will see you tomorrow.
