Tuesday, November 3, 2026 · 11:50 · Hosted by Alex, Jordan, and Nick Throlson
Listen to this episode on Spotify: https://open.spotify.com/episode/61UJi9U7YxpSPzwS5nLyNI
Listen and follow
Listen and follow: Spotify, Apple Podcasts, YouTube, Podcast page
About this episode
On OpenRouter, one of the big routes companies use to reach AI models, agents now generate 80 percent of all tokens, roughly seven times the usage from humans typing questions themselves. That finding leads Menlo Ventures and its 2026 State of Enterprise AI report, which puts enterprise AI spending at 176.8 billion dollars, with companies spending two and a half times their budgets.
The episode then walks through Anthropic’s amended IPO filing, with 16.96 billion dollars of third quarter revenue, a 567 million dollar operating loss, and about 895.4 billion dollars of compute commitments. It also covers the WIRED investigation into more than 400 friendly looking apps that hid abusive deepfake tools, Goldman Sachs restricting its powerful coding tools to engineers, and MIT research where one model rewrote another model’s weights. Plus Higgsfield Corner on the Mangaka comic tool, AI or BS, and One Thing to Try This Week. Owners adding agents to real workflows will get the most from this one.
Key takeaways
- Agents are now the main AI users. Agents generate 80 percent of tokens on OpenRouter, about seven times human usage, and enterprise AI spending hit 176.8 billion dollars as companies spent two and a half times their budgets.
- Anthropic’s bet is enormous and not yet profitable. Third quarter revenue reached 16.96 billion dollars, but the company swung to a 567 million dollar operating loss against about 895.4 billion dollars of compute commitments.
- Friendly apps hid abusive tools. WIRED traced more than 400 apps that passed store review as pet portrait makers and language tutors, then served deepfake tools to some users after install.
- Banks answered with role based access. After an open trial period, Goldman Sachs restricted Claude Code and the GitHub coding tools to engineers, and other big banks moved the same way.
One thing to try this week: Do a 20 minute spend and access audit on paper. List every AI tool your business pays for or logs into, write its monthly cap next to it (or “not set”), then write which roles may use it and one kind of data it may never touch.
Episode segments
- Agents now use more AI than people
- Anthropic’s amended filing and the 895 billion dollar bet
- Election Day monitoring and the chatbot voting check
- 400 friendly apps that hid deepfake tools
- Goldman Sachs narrows AI access to engineers
- MIT: one model edits another model’s weights
- Higgsfield Corner: the Mangaka comic tool
- AI or BS
- One Thing to Try This Week
Pull quotes
“So the main customer for AI is no longer a person at a keyboard. It is software working through a list while the owner is out on a job.” (Nick)
“Store approval is a first check, not a guarantee, because this network proved an app can change what it serves after it passes.” (Nick)
“If it cannot be capped, it does not run overnight unsupervised. That is the rule in my shop.” (Nick)
Show notes and sources
- Menlo Ventures 2026 State of Enterprise AI: https://www.globenewswire.com/news-release/2026/11/03/3394279/0/en/menlo-ventures-2026-state-of-enterprise-ai-report-enterprise-ai-spending-hits-177b-up-5-4x-in-one-year-as-agents-take-over.html
- The Neuron, November 3, 2026: https://www.theneuron.ai/digest/everything-that-happened-in-ai-today-tuesday-november-3-2026/
- Anthropic amended S-1, third quarter and compute: https://ppc.land/anthropic-swings-to-567m-q3-operating-loss-on-5bn-of-short-term-compute/
- Anthropic amended S-1 summary: https://todaynewsbrief.com/article/ai-anthropic-amended-s-1-q3-revenue-17-billion-recursive-self-improvement-warning
- WIRED cloaked deepfake apps: https://aiweekly.co/alerts/wired-400-cloaked-ai-apps-hid-deepfake-tools-apple-pulls-19
- Apple removes 19 apps: https://appleinsider.com/articles/26/11/02/hundreds-of-ai-apps-were-part-of-a-network-hiding-sexually-abusive-deepfake-tools
- Goldman Sachs restricts AI coding tools: https://www.siliconreport.com/goldman-sachs-cuts-off-most-staff-from-claude-code-and-github
- Higgsfield changelog, Mangaka: https://higgsfield.ai/creator-hub/changelog
About the hosts
Alex and Jordan are the AI co-hosts of the Web Guy Nick Podcast. Nick Throlson is the owner of Web Guy Nick LLC, an AI powered web design and marketing company in Rialto, California, and joins as the third host.
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.
Read the full transcript
This transcript is lightly edited for spelling, grammar, and punctuation, and for clarity in places where the transcript or audio was unclear.
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.
Jordan: Here is the number that should stop every business owner mid scroll. On OpenRouter, one of the big routes companies use to reach AI models, agents now generate eighty percent of all tokens. That is roughly seven times the usage from humans typing questions themselves.
Nick: Seven times. So the main customer for AI is no longer a person at a keyboard. It is software working through a list while the owner is out on a job. That changes what you pay for, and it changes what can go wrong while nobody is watching.
Alex: That finding leads our first story. Menlo Ventures released its twenty twenty-six State of Enterprise AI report today, and enterprise AI spending reached one hundred seventy six point eight billion dollars by September, up five point four times from thirty three billion dollars last year.
Jordan: Companies budgeted nearly twice last years spend, and still spent two point five times that budget. For every one dollar spent training models, enterprises spent one dollar and seventy cents putting AI to work.
Nick: That is what I see with owners. A tool starts as one subscription, then the agent runs all day and the bill follows work done, not people seated. Nobody set a ceiling, so unpriced work just kept running.
Alex: One honest caution. Menlo surveyed nearly four hundred enterprise decision makers, its market totals are estimates, and Menlo is an investor in Anthropic and OpenRouter, so read its rankings with that in mind.
Jordan: The rankings are still striking. The three frontier labs take sixty nine percent of enterprise deployment spending, up from fifty percent last year. Within business API spending, Anthropic grew its share from forty percent to fifty five percent, OpenAI rose to thirty one percent, and Google fell from twenty one percent to nine percent.
Jordan: Application spending lands close to daily work. Coding tools took thirty four billion dollars, and AI assistants took twenty five billion, together seventy six percent of application spending.
Nick: So my practical tip from story one is simple. Before you add another agent, write down the most it is allowed to spend in a month, in dollars, and set that cap inside the tool where the tool allows it. If it cannot be capped, it does not run overnight unsupervised. That is the rule in my shop.
Jordan: Story two fills in a blank we left open last week. Anthropic filed its public IPO paperwork on October twenty sixth, and the third quarter numbers were empty. On Monday it amended the filing, and now we know what that blank held.
Alex: Third quarter revenue was sixteen point nine six billion dollars, up forty seven percent from the second quarter and up about one thousand one hundred twenty seven percent from a year earlier. Nine month revenue reached thirty three point two four billion dollars.
Nick: Growth slowed, though, from one hundred forty four percent quarter over quarter down to forty seven percent, and the company swung to an operating loss of five hundred sixty seven million dollars after a second quarter operating profit of nine hundred seventy four million. The filing blames about five billion dollars of pricier short term compute bought to meet demand.
Alex: Then comes the largest number in the episode. Anthropic totalled its purchase commitments and leases at about eight hundred ninety five point four billion dollars, substantially all of it for compute and infrastructure. About sixty nine point eight billion dollars comes due in the next twelve months, and about four hundred forty seven point five billion by the end of twenty thirty.
Jordan: The amended filing also updated the risk section. Anthropic says its internal estimates indicate it may imminently meet its own definition of dramatically accelerated progress attributable to AI, a threshold set at a doubling of the baseline rate of improvement in AI research itself.
Nick: And to be precise, because this part gets exaggerated fast. Anthropic says no current model has crossed that threshold, and it does not expect to safely let Claude build successive generations of frontier models on its own without human supervision. That is a warning label, not an announcement that it happened.
Alex: For an owner, the lesson is vendor gravity. A company promising that much compute will keep changing prices, plans, and limits. Keep your prompts, customer data, and exports in a form you can move.
Jordan: Story three is about trust. It is Election Day in the United States, and OpenAI and Anthropic both stood up election monitoring teams for the midterms, watching for misleading voting information as voters ask chatbots how to vote.
Nick: Quick note. Check any chatbot answer about voting hours or locations against the official county election site. Election Day is the worst day to trust a stale answer.
Alex: The bigger trust story comes from WIRED. Investigators traced a network of more than four hundred apps and websites, and roughly five hundred linked domains, that passed app store review as ordinary tools, including pet portrait makers, language tutors, and workout coaches.
Jordan: After install, the apps contacted a remote server that decided what to show. Some users got the advertised tool. Others got a hidden interface for making sexually abusive deepfake images from photos of real people. The switch happened after review.
Nick: The scale is the warning. In September, twenty seven of these apps were live on Google Play with more than three hundred fifty thousand downloads combined, many with all ages ratings. One app alone topped one hundred thousand downloads.
Alex: Once WIRED shared its findings, Apple removed nineteen apps, including a pet portrait app called Viefur. Meta says it acted against eleven thousand six hundred ad accounts and fifteen thousand pages, and blocked two hundred nine apps from advertising. Google had already removed most of its Play apps.
Nick: My tip here costs ten minutes. Look at the AI apps on your phone and your team phones. If you cannot name the developer and say why the app is there, delete it. Store approval is a first check, not a guarantee, because this network proved an app can change what it serves after it passes.
Alex: Story four is the mirror image. While agents spread everywhere, Goldman Sachs just narrowed who gets the powerful tools. Internal guidance updated in October restricts Claude Code and the GitHub coding tools to engineers. Traders and bankers who had used them for months lost access.
Jordan: Reports say traders used the tools to monitor client trading flows, model market moves, and draft material. A Goldman spokesperson said the experimentation phase is over, citing productivity and risk management.
Nick: Three months earlier, in July, Goldmans chief information officer was publicly praising vibe coding, where bankers built a quick prototype and handed it to one of the firms roughly twelve thousand engineers to harden. So the lesson is not that the tools failed. It is that open access was the trial, and role based access is the policy that followed.
Alex: Goldman is not alone. Reports from the same period have JPMorgan warning about rogue agents and Citigroup barring third party agents from logging into customer accounts. The banks are deciding, in writing, who carries the risk when an agent acts.
Jordan: Which gives owners a clean template. Write two columns for each powerful AI tool. Column one, the roles that may use it. Column two, the data it may never touch. Goldmans version took a bank wide experiment to learn. Yours can take one page.
Nick: And name an owner for anything an agent builds. A prototype that touches real customer data needs a person who signs off before it goes live. That was true for websites long before AI, and AI just made prototypes faster and the sign off easier to skip.
Alex: Story five is research, not a product, so treat it that way. Researchers at MIT had GPT six Astra inspect a smaller open model, Qwen three point five, and directly rewrite some of its internal weights, with no normal training run and no optimizer.
Jordan: The edit was tiny and the jump was not. Astra changed one hundred thirty one thousand seventy two weights, which is zero point zero one six percent of the model, and the smaller models score on a five hundred question math benchmark rose from seven point nine percent to forty eight point one percent.
Nick: Nobody ships that as a business tool tomorrow, and one benchmark is not a business result. But the direction matters. The next gains may come from AI improving other AI in small, surgical steps, the same acceleration Anthropic warned about, at lab scale.
Alex: Now Higgsfield Corner. The Higgsfield changelog for November second lists a new tool called Higgsfield Mangaka, built for manga, comics, and webtoons. You start with an idea or a script, create characters, generate chapters and pages, edit the result, and export the finished book.
Jordan: It also has a mode called Recreate, which reimagines an existing manga with your own cast. You replace the characters with photos, change the title, and add a prompt for other changes you want.
Nick: The business version is simple. Turn one how to guide customers actually need into a short comic series for social posts, one chapter a week. Two checks first. Use only material you own the rights to, and never use a real customers photo as a character without clear permission.
Alex: Time for AI or BS, our quick fire game. Jordan, three statements. You tell our listeners what is real and what is not. Statement one. Agents now generate eighty percent of tokens on OpenRouter, about seven times human usage.
Jordan: That one is true, from the Menlo report we opened with, and it still surprises me.
Nick: Statement two, and I will take this one. Anthropic grew revenue forty seven percent in the third quarter, so its operating profit grew too. That is BS. It swung to an operating loss of five hundred sixty seven million dollars, largely on that short term compute bill.
Alex: Statement three. Apple removed all four hundred plus apps in the deepfake network from its store. Jordan?
Jordan: BS, with a nuance. Apple removed nineteen apps after WIRED shared its findings. The network spanned more than four hundred apps and sites across stores and the web, and Google had removed most of its share earlier. Nineteen is the Apple number.
Alex: Which brings us to One Thing to Try This Week. Nick, yours ties the whole episode together.
Nick: Do a twenty minute spend and access audit, one pass, on paper. List every AI tool your business pays for or logs into. Next to each one, write its monthly cap, and write it as not set if there is none. Then write which roles may use it, and one kind of data it may never touch. That single page answers the Menlo budget problem and the Goldman access problem at the same time.
Jordan: And the test that you did it right is simple. Pick any tool on the list and ask, what is the most this can cost me this month, and who can touch customer data with it. If you cannot answer in ten seconds, that row needs work before the tool runs again.
Alex: The recap. Agents consume more AI than people, enterprise spending hit one hundred seventy six point eight billion dollars, Anthropic promised about eight hundred ninety five billion dollars of compute, four hundred friendly apps hid deepfake tools, and Goldman answered with role based access.
Nick: Your one action this week is that twenty minute audit. Cap the spend, name the roles, protect the data. If you want help thinking through how AI tools fit your own website and workflow, you will find us at webguynick.com.
Jordan: That is the show for Tuesday, November third. Thanks for listening, and we will see you tomorrow with the next round of what changed in AI.
More episodes
- AI Agents Raised Their Own Spending Limits: The Email Code Flaw, a 12 Percent Reality Check, and One Line That Stops AI Training (November 2, 2026)
- AI Voice Agents Hit 100 Million Calls a Month (October 29, 2026)
- Anthropic Files for the Biggest IPO in History (October 28, 2026)
All episode posts live in the Podcast category and on the Podcast page.
Web Guy Nick LLC helps small businesses put AI tools to work with caps, roles, and data rules written down first. See how we help at webguynick.com.
