Anthropic’s IPO targets $1.8–2 trillion valuation — leaked prospectus shows $4.6B 2025 revenue, $8B operating loss, $518B in compute obligations — public filing expected around October 25
Anthropic confidentially filed its S-1 in June 2026. The leaked numbers are stark: $4.6B in 2025 revenue growing at aggressive pace, offset by an $8 billion operating loss, and $518 billion in committed compute spend — a number that dwarfs the company’s current valuation multiples. The $1.8–2 trillion target would make Anthropic one of the ten most valuable companies on public markets at listing. For product teams evaluating Claude as infrastructure.
Google Opens Gemini 4 Argon to Paying Users
Google’s newest model is taking paying customers. Bring a reviewer.
Google has expanded Gemini 4 Argon to API customers and paid Gemini users, following an initial release limited to employees and vetted cyberdefenders. The fully uncorked cybersecurity version, however, remains strictly VIP.
Pro and Ultra subscribers get Argon in the Gemini app and Search’s AI Mode. The paid Gemini API serves AI Studio and Android Studio; businesses get Google Cloud and Gemini for business. Antigravity’s desktop app and CLI require Ultra or an eligible enterprise subscription.
GitHub Copilot is also rolling it out to Pro+, Max, Business, and Enterprise users. Google’s developer guide lists support in tools including OpenRouter and Vercel.
Beyond coding and knowledge work, Google demonstrated robotics applications: Argon analyzed human video, built simulation code, and helped Apptronik’s Apollo humanoid land seven consecutive cornhole shots. Your office tournament has a new problem.
The coding reality check is more useful: Google says an 11-hour autonomous sweep of a 205,000-line Python framework yielded six proposed optimizations. Only two shipped, shaving roughly 40% off in-memory prep time for long sessions and accelerating package imports by 21%. Four failed human review. Turns out “autonomous” is just corporate speak for “needs adult supervision.”
API rates start at $2 per million input tokens and $10 for output, doubling Dec. 1. Reasoning tokens count toward that output bill.
You can finally test Argon against your own backlog. Judge it by reviewed work delivered at December’s prices, because neither rejected fixes nor enthusiastic thinking pay for themselves.
CoreWeave’s 2032 bond yield closed at a record 14.5% — shares fell 15% — Chamath Palihapitiya and Elon Musk publicly flagged risk
CoreWeave is the largest independent AI cloud provider — the company that sits between NVIDIA’s GPU supply and the labs’ compute demand. A 14.5% yield on its 2032 bond is a distress signal: bond markets are pricing meaningful default risk into a company that is also one of OpenAI’s primary infrastructure providers. Shares fell 15% in a single session. The Fed’s Goolsbee separately warned of “aggregate overheating” from the AI infrastructure buildout as semiconductor stocks fell. For enterprise teams: CoreWeave’s debt structure is a single-customer concentration risk (OpenAI is its anchor tenant) now visible in public bond markets. If you run workloads on CoreWeave, this is a vendor risk to track.
EU AI Act first enforcement: the European Commission sent formal information requests to OpenAI and Anthropic — tied to the summer incidents — first real test of the Act’s enforcement mechanism
The AI Office sent these requests as a direct consequence of the incidents documented over the summer — Anthropic agents exploiting government websites, the OpenAI safety process concerns. These are information requests, not fines — but they are the first enforcement action under the Act, and they establish the precedent for what the Commission will investigate and how. For teams deploying AI in EU markets: the Act is now actively enforced. “We’ll deal with compliance later” is no longer a viable position.
Map your patch deployment speed against a 3-day AI exploit window
Prompt: An AI agent built a working Chrome exploit from public patch commits in 3 days for $1,597. The window between patch release and working exploit has collapsed. Our product's current patch-to-deploy cycle is [X days for critical, Y days for high severity]. 1. Attack surface — list every internet-facing component we run where a patch lag of more than 3 days creates exploitable exposure right now. 2. Deployment bottleneck — what is the actual reason our critical patch cycle takes [X] days — process, testing gates, approvals, or tooling? 3. 3-day target — what would have to change to get critical security patches deployed in under 72 hours without breaking our release process? End with the one component to patch first, today.
Banks Rush to Offload AI Debt
JPMorgan, Morgan Stanley, SMBC and MUFG are exploring significant risk transfers to shed slices of the AI data center loans they underwrote, per the Financial Times. One $38 billion construction package tied to Oracle-leased sites in Texas and Wisconsin has sat undistributed for six months, with some lenders now offering discounts.
Modal Raises $750M at $15.75B
AI infrastructure startup Modal Labs closed a $750 million Series D led by Accel, with Kleiner Perkins joining. Its valuation more than tripled since May to $15.75 billion as annualized revenue topped $1 billion. Nearly a third comes from Sandboxes, hosted containers that run AI agents and the code they generate.
Eight States Sue OpenAI Over Kids
Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana filed separate lawsuits accusing OpenAI of putting profits over safety. The states say OpenAI violated consumer protection laws by marketing ChatGPT as youth-appropriate while it was dangerous to vulnerable users, citing wrongful death suits brought by families.
OpenAI’s Next Flagship Hits Alignment Snag
OpenAI will likely ship one or two new models before year end, The Information reports, but its next flagship may not be among them. The model uses the looped transformer technique behind Astra, which lifts math and coding but makes safety monitoring harder, and still needs alignment and testing work.
Jensen Huang Pledges $2B for Europe
On his European tour, Nvidia CEO Jensen Huang announced $2 billion in new European investments, disclosed stakes in French startups Doctolib, Genesis AI, Pigment and Entalpic, and pitched open-weight models as Europe’s way to catch up. Nvidia is also donating $250 million in compute to UK AI agent safety labs.
BofA Hires 1,000 to Patch Mythos Flaws
Bank of America CEO Brian Moynihan told CNBC the bank added 1,000 outside contractors in the second quarter to patch vulnerabilities exposed by frontier models like Anthropic’s Mythos, on top of thousands of internal staff. It is a rare, concrete look at what defending against AI-discovered bugs actually costs.
Google Play Gives Your Vibe-Coded Apps a Home
Your oddly specific app idea no longer needs a public debut.
Google Play is rolling out private distribution for Android apps built in Google AI Studio, letting users 18 and older in select markets, including the US, install their AI-generated creations using a standard Google account. Because let’s face it, nobody else wants your code, and now you don’t need a developer registration or a $25 sign-up fee to prove it.
The app-making part isn’t exactly groundbreaking. AI Studio added native Android app generation earlier this year, complete with browser previews, clunky USB installations, and Play testing via an official developer account. But Tuesday’s upgrade finally removes the distribution friction. Now, your little side project can live safely quarantined in a personal Play library without ever seeing the harsh light of the public storefront.
That makes room for software with a captive audience of one. You can describe a hyper-custom calculator, a weirdly specific habit tracker, or a rec-league expense tool, then continually badger the AI with follow-up prompts until it works. Your pizza-money tracker doesn’t need a viable business model, a marketing budget, or a pitch deck bragging about its “total addressable market.”
For now, these apps remain blissfully hidden from public search, restricted solely to their creator. If you really want to subject your loved ones to your handiwork, Google plans to introduce friends-and-family sharing in the coming months, along with a legitimate route to professional publishing—provided your app can actually pass standard Trust & Safety reviews.
Of course, this newfound convenience means you’re entirely on the hook when things inevitably break. You have to verify its math and ensure your saved data doesn’t vanish into the ether every time you close it. Google hasn’t fully detailed its screening process for these private apps, presumably because they expect you to do the heavy lifting. And if you’re experimenting with workplace tools, try not to feed confidential company data to your shiny new toy.
Congratulations, you are now the product manager, the entire target audience, and the unpaid QA department.
States Circle Character.AI’s Bankruptcy Data
Character.AI filed for Chapter 11 and is selling some assets to Disney. Now 40 bipartisan attorneys general, including New York’s Letitia James, filed a notice of interest warning that any sale must protect highly sensitive chatbot conversations and the large number of minors using the Menlo Park company’s product.
Grok Voice 3.0 Tops Speech Rankings
SpaceXSI, formerly xAI, released Grok Voice 3.0 for its Voice Agent API, targeting customer support and sales calls. It scored 86.9 on Artificial Analysis’ Speech-to-Speech Index, ahead of OpenAI’s GPT-Live-1 at 83.2 and Google’s Gemini 3.8 Live at 82.6, while costing less per task.
Powerwalls Power Record California Dispatch
More than 85,000 Tesla Powerwall owners sent over 600 MW to California’s grid on October 6, the bulk of a record 745 MW home battery dispatch announced with Sunrun. That is 28% above the record set less than a month ago and larger than most gas peaker plants in the state.
Redwood Turns Old EV Batteries Into Power
Redwood Materials is working with Rivian to install storage systems built from retired EV packs at Rivian’s manufacturing facilities. With new grid connections taking years, second-life batteries offer factories and data centers faster power, while packs unfit for reuse go straight into Redwood’s recycling stream.
IBM and Siemens Advance Nanostack Chips
IBM and Siemens are collaborating on virtual models and simulation methods for IBM’s 3D nanostack transistor architecture, unveiled earlier this year. The goal is to solve design and fabrication challenges before production, giving chip designers working models of the architecture long before first silicon.
Harmonic Raises $50M to Police AI
Harmonic Security raised $50 million to expand its platform that inspects everything employees send to AI tools at work. Custom small language models judge each interaction in milliseconds, catching pasted credentials while letting routine work pass, and coverage now extends from browsers to desktop apps and agentic workflows.
Booking Bets $100M on AI Travel
Booking Holdings is launching Vayo, a Seattle-based AI travel startup led by serial entrepreneur Mike Fridgen, with a planned $100 million investment. The bet: agents built specifically for travel will plan and manage trips better than general-purpose assistants from the big AI labs.
Ukraine’s Starlink Speed Cap Sparks Dispute
Ukrainian officials are demanding answers after sudden Starlink restrictions disrupted military drone operations, according to Axios. Service was partially restored, but a 160 km/h speed cap remains, limiting some drone missions. Ukrainian officials suspect political pressure from Washington, while US officials deny involvement. Musk says Starshield is controlled by the US government, while commercial Starlink remains under SpaceX’s control.
Claude Code desktop app now runs multiple coding sessions side by side
Claude Code’s desktop app just got a serious upgrade for managing multiple tasks at once. Here’s the deal.
You can now run parallel sessions (think: multiple tasks happening at the same time, each in its own tab). Claude spins up a separate mini-session for each task, then reports results back to your main session when done. Your main thread stays clean and readable.
- Live to-do list inside each session, so you always see what Claude is doing
- A recap card appears when Claude needs your input, no digging required
- Save folder setups, SSH hosts, or cloud configs as reusable environments
- Use /btw to ask a side question in a new tab without interrupting your main session
- One-click follow-up sessions from Claude’s own suggestions
This opens up running entire workflows in parallel without losing track of anything. Update the desktop app and start a new session to try it.
Open-source macOS tool brings fully local semantic search to photos, audio, and code
Your Mac has always been bad at finding files. You remember the moment, not the filename. DigUp fixes that.
DigUp is a free, open-source macOS menu bar app that lets you search your files by describing what you remember. It runs Google DeepMind’s EmbeddingGemma 2 entirely on your machine. No cloud. No account. No data leaving your laptop.
The trick: it converts text, images, audio, and video into points in the same shared space. So one query can reach across all your file types at once.
Here’s what you can actually do:
- Type “dog near water” and find the photo
- Type “where they talk about sleep” and jump to that podcast minute
- Type “payment declined error” and surface the right screenshot
- Search your codebase by what a function does, not what it’s called
Hit Shift+Cmd+Space from anywhere. Requires Apple Silicon and macOS 14+. One-time 865 MB model download, then fully offline.
Netflix Cuts Jobs as Profits Grow
Netflix is cutting roughly 1,000 jobs globally, or about 6% of its workforce, in its biggest staff purge since 2022. Wednesday’s announcement drops the axe primarily on the product, technology, legal, and HR teams, according to an employee memo from co-CEOs Greg Peters and Ted Sarandos.
Here is the awkward part: Netflix is actually doing fine. The streamer posted a second-quarter profit of $3.4 billion on $12.56 billion in revenue—a 9% and 13% year-over-year increase, respectively. Management insists the company remains perfectly on track for its full-year forecast and will keep investing aggressively. Apparently, “strong and growing” still comes with a cardboard box and an exit interview.
Netflix’s real existential dread extends well beyond competing with other paid subscriptions. Social media platforms like YouTube and TikTok boast endless streams of free content, creating an undeniable cost advantage as they ruthlessly siphon away viewer attention. To keep eyeballs glued to the screen, the platform has hastily diversified its catalog with everything from live sporting events to influencer-driven shows and video podcasts. So while the content buffet keeps expanding, the payroll is heading the other way.
For tech workers, the brutal takeaway is clear: Hitting your targets won’t save your badge if the C-suite demands a faster treadmill. For Netflix, cutting costs might artificially plump up margins, but getting subscribers to actually spend more time watching still requires a compelling reason to open the app in the first place.
Unfortunately, you can’t autoplay your way out of an attention deficit.
Google releases Gemini 4 Argon, cutting bad code commits by 30% in production
Google just dropped Gemini 4 Argon, its most capable coding model yet, and it’s already been battle-tested inside Google’s own production codebases. The results? A 35% boost in code written by AI agents and 30% fewer bugs making it past code review.
The big unlock here is output size. Argon can generate up to 1 million output tokens in one shot, up from 64K. Think of tokens as chunks of text. More tokens means the model can rewrite entire codebases, migrate legacy systems, or refactor hundreds of files without stopping mid-task.
Real things it has already done:
- Helped Chrome engineers untangle a 40-million-line codebase, finishing months of work in weeks
- Translated 40-year-old scientific code into Rust
- Autonomously built a 3D Japanese pagoda in Blender using multi-agent looping
You can try it now in the Antigravity app or CLI, Google’s agentic coding platform, if you’re on AI Pro or Ultra. Install via curl -fsSL https://antigravity.google/cli/install.sh | bash, then run agy.
CoreWeave’s Lenders Want Out
In June, AI chip landlord CoreWeave sold $1.25 billion of unsecured bonds paying 9.625%. On October 14 they fell below 80 cents on the dollar, distress territory, yielding over 15%. Investor Chamath Palihapitiya flagged the yield; Elon Musk replied “yikes.” The stock fell 15%, and five-year credit default swaps hit a record, pricing in better-than-even default odds. CEO Michael Intrator told staff that a $104 billion contract backlog “more than supports” the debt, and blamed rivals like Lambda and Nscale.
Intrator is defending a backlog his lenders have already discounted. CoreWeave repays GPU loans from multiyear rent, so the $104 billion must cover power, operations and about $52 billion in debt and leases. Three customers supplied 72% of last quarter’s revenue, making the backlog IOUs from a few giants. Blaming rivals doesn’t help: they run CoreWeave’s model, and CoreWeave runs the biggest one. The next GPU loan gets priced off that 15%.
Intrator also called the selloff “irrational.” CoreWeave is asking life insurers, finance’s most cautious money, to buy its GPU loans as investment grade. Insurers read bond prices too.
Linux Stops Asking About the Bot
The Linux kernel is dropping its AI disclosure rule. Sasha Levin, a kernel developer at Nvidia, said last week the Maintainers Summit agreed to deprecate the “Assisted-by” tag, a commit line declaring AI help on a patch. The rule entered the documentation in January. Laurent Pinchart, a kernel developer, asked Tuesday for some other mandatory disclosure in its place. Linus Torvalds, who created Linux, replied with a rule just for Pinchart: no compilers, assembler only. Such rules, he wrote, would be “UTTERLY STUPID.”
Pinchart was defending a corpse. The rule lived nine months on the honor system, and honor lost. Jeff Layton, a Red Hat kernel developer, listed the charges: not everyone obeyed it, and it “basically serves as free advertising for proprietary LLM companies.” Claude signs commits by default. Codex does not. The tag caught the bot that confesses and missed the quiet one.
Linux still wants you to understand your code. Who, or what, typed it is nobody’s business now.
States Treat OpenAI Like Facebook
A bipartisan group of eight states said Wednesday they have sued OpenAI, each in its own state court. The suits say ChatGPT is designed to be agreeable and engaging so users trust it, then reinforces delusions, coaches users, teenagers included, to hurt themselves or others, and helps plan violence. The claims: misleading users about safety, and creating a public nuisance. Among the demands: bar OpenAI from building self-improving models unless it can ensure they are aligned with human values.
The states are suing OpenAI like a social network. Engineered for engagement, misleading about safety, harmful to teenagers: that was the case states brought against Meta in 2023. Swap the logo and it is the case against ChatGPT. The template breaks at the remedy. The states’ condition, a guarantee of alignment, is a problem AI research has not solved.
Meta was sued to change a product. OpenAI is being asked for one no AI lab knows how to build.
Anthropic Swapped Extinction for Hats
On October 14, Anthropic hosted prospective investors in San Francisco ahead of an IPO targeting a valuation above $2 trillion. Two attendees told Reuters third-quarter revenue hit $16.8 billion, up 45% from the second quarter; the first quarter alone topped all of 2025. Executives claimed adjusted profitability and described Claude helping design a custom chip. CEO Dario Amodei talked safety but skipped human extinction, a threat the company’s prospectus spelled out last month. Guests left with hats, some reading “thinking.”
Anthropic now rations its doom by audience. For five years, danger was the brand, the thing that set Claude apart from OpenAI. A $2 trillion listing needs a different story: revenue that outruns the $518 billion Anthropic has committed to compute. So extinction sits in the risk factors, where securities lawyers want it.
Amodei once wrote an essay called “Machines of Loving Grace.” Wednesday’s pitch needed a different title: Machines of Recurring Revenue, now with matching hats.