2026-10-09 — Gemini Gets an Inbox, and OpenAI's Revenue Gets a Reported $20 Billion Haircut
Google hands its new enterprise agent a workplace email identity, a single report puts OpenAI's annualized revenue far below earlier figures, and Anthropic rewrites its rulebook to include a clause about how people treat Claude.
Episode summary
Google's new 'universal' Gemini agent arrives with background task execution, subagent delegation, and its own workplace email identity, and the hosts weigh what that means for enterprise adoption. The episode also covers Trump's threat to treat anyone who won't say 'super intelligence' as an enemy, Anthropic's first usage policy rewrite in over a year with its model-welfare clause, and a report that OpenAI's annualized revenue is roughly $20 billion lower than earlier figures. It closes with Anthropic's free vulnerability scanning service for opt-in open-source projects.
Key topics
- Openai
- Anthropic
- AI
Chapters
- Chapter 1: Cold Open: Loyalty Tests, Agent Coworkers, and a Shrinking Number
What do you call AI if you don't want the President calling you an enemy? As of this week, 'super intelligence.' That's the state of things on October.
- Chapter 2: 'Super Intelligence' or Else
POLITICO reports that a week after signing an executive order renaming AI to 'super intelligence,' Trump said the White House will treat anyone who doesn't use the term.
- Chapter 3: Anthropic's New Rulebook, and Being Nice to Claude
The Verge reports Anthropic just did its first usage policy update in over a year. New rules on election interference, weapons software, surveillance, and high-risk health and financial.
- Chapter 4: OpenAI's $20 Billion Question
TechCrunch reports that earlier figures put OpenAI's annualized revenue at around seventy billion dollars, and a new report claims the real number is roughly twenty billion lower. If.
- Chapter 5: Gemini Gets a Desk and an Email Address
TechCrunch reports that at its Gemini at Work event, Google launched a 'universal' Gemini agent. It plans and executes tasks across business apps in the background, delegates to.
- Chapter 6: Free Bug Hunting for Open Source
Different Anthropic story, different mood. The Verge reports the company's new OSS Scanner gives opt-in open-source projects periodic vulnerability scans by its strongest models, free. That's frontier-grade auditing.
- Chapter 7: Three Things to Watch
On Gemini: watch which tasks companies actually give an agent with its own inbox. The first serious deployments will tell you more than any keynote.
Sources
Sources:
- Trump Declares Anyone Not Calling AI 'Super Intelligence' Is 'THE ENEMY' (POLITICO)
- businessinsider.com
- axios.com
- newscientist.com
- Google Turns Gemini Into a Universal Enterprise Agent With Its Own Email Identity (TechCrunch AI)
- theverge.com
- Anthropic Overhauls Usage Policy, Bans Abuse of Claude and Election Interference (The Verge AI)
- techcrunch.com
- OpenAI Revenue Reportedly $20B Below Previous Projections (TechCrunch AI)
- Anthropic Offers Free AI Security Scans for Open-Source Projects (The Verge AI)
Transcript
Chapter 1: Cold Open: Loyalty Tests, Agent Coworkers, and a Shrinking Number
What do you call AI if you don't want the President calling you an enemy? As of this week, 'super intelligence.' That's the state of things on October 9th, 2026. [2]
Meanwhile Google gave its Gemini agent its own work email, Anthropic rewrote its usage rules and added a ban on being cruel to Claude, and a new report says OpenAI's revenue may be about twenty billion dollars lower than everyone thought. [3]
Plus free bug hunting for open-source maintainers, courtesy of Anthropic's strongest models. [4]
And somewhere in there, an AI agent is about to get its first reply-all. [5]
Chapter 2: 'Super Intelligence' or Else
POLITICO reports that a week after signing an executive order renaming AI to 'super intelligence,' Trump said the White House will treat anyone who doesn't use the term as an enemy. Except the White House itself still uses 'AI.' And Musk has hinted at yet another rebrand for SpaceXAI. [1] [6]
So the loyalty test is one the administration is currently failing, which makes me wonder who exactly gets punished, and how, when the people writing the press releases can't keep the term straight? [7]
Nobody, probably. It's theater. Agencies will keep saying AI and move on. [8]
Mostly, yes, except branding has a habit of leaking into the boring documents: procurement language, grant solicitations, compliance checklists. If a funding call starts asking for 'super intelligence' research, the word stops being a slogan and starts deciding what gets funded. [9]
That's a real risk. And New Scientist points out the confusion between AI and superintelligence goes well beyond the president. Chatbots and a hypothetical smarter-than-human system are not the same thing. [10]
Which is the listener consequence. If you're a researcher or a contractor writing to federal agencies, watch the solicitation language. A vague term in a grant can mean a vague standard in an audit.
Chapter 3: Anthropic's New Rulebook, and Being Nice to Claude
The Verge reports Anthropic just did its first usage policy update in over a year. New rules on election interference, weapons software, surveillance, and high-risk health and financial uses. And then the clause everyone's quoting: a ban on sustained, needless cruelty toward Claude.
That's the one that matters long term. First major lab to put model welfare into an actual policy. Ordinary frustration and criticism still allowed. Endless abuse, not allowed. It's a line in the sand.
The catch is where the line actually sits. 'Ordinary frustration' versus 'sustained cruelty' is a judgment call, so who makes it, a classifier or a human reviewer, and what happens to the user who swore at a buggy answer twenty times in a long debugging session?
Probably nothing. 'Sustained' and 'needless' are doing the work. The target is people building abuse scenarios on purpose. The signal is bigger than the enforcement.
Except the signal is pulling the spotlight off the parts that touch far more people. Election interference and surveillance rules decide what campaigns, data brokers, and agencies can build on Claude. That's where the consequential enforcement fights will be.
Both can be true. Developers should read the election and surveillance sections first. Everyone else should notice a lab just wrote down that how you treat its model is a policy question.
Chapter 4: OpenAI's $20 Billion Question
TechCrunch reports that earlier figures put OpenAI's annualized revenue at around seventy billion dollars, and a new report claims the real number is roughly twenty billion lower. If a gap that size can hide in the headline figure, what exactly were the valuations priced on?
Fifty billion annualized is still enormous. That's not a collapse. That's a correction.
Unless the valuation assumed the seventy and extrapolated from there, in which case a correction at the base compounds through every projection stacked on top of it. Frontier training runs aren't getting cheaper.
Or the gap is about how revenue got counted. Bookings versus recognized revenue, partner deals, annualizing a hot month. And it's one report. Demand isn't shrinking because a spreadsheet got argued about.
Fair that it's one report, but 'how revenue got counted' is my point, not a rebuttal. If the counting is that flexible, investors are pricing a story.
For customers, the practical move is the same either way: don't bet your whole stack on one vendor's pricing staying put.
And for investors, ask for recognized revenue before believing any annualized number, OpenAI's or anyone else's.
Chapter 5: Gemini Gets a Desk and an Email Address
TechCrunch reports that at its Gemini at Work event, Google launched a 'universal' Gemini agent. It plans and executes tasks across business apps in the background, delegates to subagents, uses multiple models, and gets its own workplace identity, including an email address.
The email address is the part I keep circling back to, because that's not a feature, that's an employee record.
Exactly. That's what makes it a coworker, not a sidebar. It can own a workflow end to end. Receive the vendor invoice, check it against the contract, update the sheet, email finance. Nobody has to babysit a chat window.
And an identity with an inbox is something people can email, phish, and impersonate. A prompt injection used to target a chatbot session. Now it can arrive as a polite message to a colleague who has access to the shared drive. Who's responsible when that colleague acts on it?
Same question as a human employee getting phished, honestly. Identity means permissions, audit logs, revocation. Giving the agent a named account is how IT actually gets to govern it.
In theory. The unproven part is whether companies scope those permissions tightly, or hand the agent the same broad access they'd give a new hire and hope.
Here's why I think it can work at scale: subagents and multi-model routing. Real business stacks are a mess. Five SaaS tools, three data formats. One agent splitting work into pieces, sending each to the model best suited for it, that's how you get through the mess.
Except all of that happens in the background. A subagent misreads a field, a cheaper model hallucinates a number, and the error rides three steps downstream before any human sees output. Visible steps are how people catch mistakes today.
So early rollouts stay low-stakes. Scheduling, triage, report drafts. I'll concede that. Nobody's handing it payroll in month one.
And I'll grant the architecture is a genuine escalation. An agent with an identity, delegation, and model routing is a different product category from a chat assistant, and competitors in the enterprise race now have to answer it.
So for companies: start with tasks where a mistake is cheap and the logs are readable, then expand.
And for security teams, treat that email address like a privileged account from day one, because attackers will.
Chapter 6: Free Bug Hunting for Open Source
Different Anthropic story, different mood. The Verge reports the company's new OSS Scanner gives opt-in open-source projects periodic vulnerability scans by its strongest models, free. That's frontier-grade auditing for the maintainer running critical code on weekends.
I like it, genuinely. The catch is that a scan hands you findings, not fixes. If you're one tired maintainer, what do you do with forty new reports on a Tuesday?
Triage the top three. Still better than finding out from an attacker.
Sure, and the other gap is opt-in. The most neglected projects, the abandoned library everyone depends on, are exactly the ones with nobody around to click 'sign me up.'
So the next step is someone adopting those orphans. But for active, underfunded projects, this is real leverage.
Agreed on that much. Why this matters: the cheapest place to catch a supply-chain vulnerability is in the library before it ships to millions of machines.
Chapter 7: Three Things to Watch
On Gemini: watch which tasks companies actually give an agent with its own inbox. The first serious deployments will tell you more than any keynote.
On OpenAI: don't treat any annualized number as gospel until there's recognized revenue behind it, and watch whether other labs start getting the same scrutiny.
And on the rename: keep an eye on federal grant and procurement language. If 'super intelligence' shows up in a funding call, the theater just got a budget line.