> daily_signal(2026_09_21)
Four fights over the AI in your life: subscribers sued the big labs, OpenAI got caught tracking you off-site, Anthropic aimed at a record IPO, and AI learned to read a cheap blood test.
PickBits Daily Signal · Monday, September 21, 2026
// tl;dr
- Four paying subscribers to ChatGPT, Claude, Grok and Gemini filed a proposed nationwide antitrust class action against Anthropic, OpenAI, xAI and Google. The complaint says the four illegally agreed to coordinate an AI slowdown after Dario Amodei's September 12 essay, and that the coordination shortchanges every subscriber. It was filed September 18 in the U.S. District Court for the Northern District of California.
- A security analysis found OpenAI's new ad pixel sets a cross-site cookie that ties your browsing to your ChatGPT account. It scrapes emails, phone numbers and names off advertiser pages, hashes those but sends your location in the clear, works when you're logged out, and keeps an anonymous ID for at least 27 days. Scraped identity events outnumbered advertiser-supplied ones 685 to 255.
- Anthropic's annualized revenue run rate jumped from about $9 billion to roughly $100 billion in a year, and it plans to go public in November. Bankers have told investors the offering could raise more than $100 billion at a valuation near $2 trillion. Sam Altman says OpenAI won't go public in 2026.
- A study in Neurocritical Care used D-dimer, a routine ER blood marker, plus machine learning to predict who recovers poorly after a brain bleed. An XGBoost model reached an AUC of 0.904 on held-out cases, and a simpler D-dimer, bleeding-volume and Hunt-Hess composite reached 0.867. It's early, single-center research, not a cleared product.
Last week we watched the biggest AI labs ask Washington for permission to slow down together, an antitrust exemption a rival called a cartel by another name. This week that fight jumped forums: four ordinary subscribers turned it into a federal class action, arguing the coordination itself is illegal and that every paying user is the one shortchanged. And the Anthropic thread we ran yesterday, about Claude leading a quarter of its own research, got a second, blunter number today: a revenue run rate racing toward a hundred billion dollars and a public offering that could be the largest in history.
The other scare today is quieter, and it's aimed right at you. OpenAI, the company that dropped "open" long ago, was caught wiring an ad tracker into the assistant hundreds of millions treat as private, following people across the web even when they never logged in. And on the brighter side, researchers showed an AI reading a blood test your ER already runs can help tell a family whether someone recovers from a brain bleed, using a marker the smallest hospital already measures.
Today: four subscribers sued Anthropic, OpenAI, xAI and Google over a coordinated AI slowdown, OpenAI's new ad pixel was caught tracking logged-out users for 27 days, Anthropic aimed at a two-trillion-dollar November IPO, and a D-dimer blood test read by AI predicted brain-bleed recovery at about ninety percent accuracy.
1. Four people who pay for AI just sued the four biggest labs for agreeing to slow it down.
The suit says the four didn't just each slow down, they agreed to, and that the coordination is what breaks the law.
Last week, the story was a request: OpenAI, Anthropic and Google were quietly asking Washington for an antitrust exemption so they could coordinate a slowdown on frontier AI, and a rival chief executive called it a cartel by another name. On Friday, September 18, that request met its mirror image in court. Four people who pay for ChatGPT, Claude, Grok and Gemini filed a proposed nationwide antitrust class action in the U.S. District Court for the Northern District of California against Anthropic, OpenAI, SpaceXAI and Google. The whole complaint hangs on one distinction: any single firm may slow its own work, but these four agreed to slow down together. It points to Anthropic's Dario Amodei publishing a September 12 essay urging industry-wide deceleration, and says Sam Altman, Elon Musk and Demis Hassabis backed it the same day. Lead counsel Nick Rowley puts it plainly: the four agreed to hold back together instead of each deciding on its own.
Skeptics were quick to say this isn't really antitrust, that none of these labs dominates a single market the way a classic monopoly does. The complaint sidesteps that entirely. It isn't alleging a monopoly, it's alleging an agreement, which is the same theory that turned landlord rent-setting software into a price-fixing case, and agreements among competitors don't need a monopoly to be unlawful. What gets me is that no public step ever happened here: no regulator approved this coordination, no proceeding weighed it, and the tens of millions of people paying for the product every month were never in the room. And the harm they point to is happening now. If the four coordinate to keep progress slower than competition would produce, the subscription you pay for is worth less than it should be.
Why this matters: You pay for one of the big AI assistants, and a new federal lawsuit says you're being overcharged for a product four rivals agreed to hold back. The case tests something bigger than a refund: whether "AI safety" can be run as a private agreement among the most powerful sellers, or whether that's just the kind of coordination the Sherman Act was written to stop. However a judge rules, it drags the slowdown fight out of essays and into a forum where evidence gets produced.
Action this week: Watch the Northern District of California docket for whether the judge certifies the class, since that's the moment this turns from a filing into a real threat. My own read is that this suit and the Washington exemption push are the same fight from opposite ends, so track both, because an exemption granted in D.C. could quietly moot the case before it's ever argued.
pbs.org (AP): Lawsuit says Anthropic, OpenAI, SpaceXAI and Google made an illegal agreement on an AI slowdown (September 19, 2026)
cbsnews.com: AI slowdown lawsuit targets OpenAI, Anthropic and Google (September 2026)
2. OpenAI's new ad tracker follows you across the web, even when you're logged out.
A security researcher found the pixel scrapes names and emails off the page and keeps an anonymous ID on you for at least twenty-seven days.
When a website runs OpenAI's new ad-measurement pixel, it quietly talks to a collector at bzr.openai.com that drops a cookie called __obi, scoped so it rides along on cross-site requests and carries a signed token binding your browsing to your ChatGPT account. A security researcher who took the system apart this weekend found it does more than use the data advertisers hand over: it scrapes emails, phone numbers and names straight off the pages it sits on. In the traffic they logged, scraped identity events outnumbered advertiser-supplied ones 685 to 255. Emails, phone numbers and names are hashed with SHA-256; your country, region, city and postal code are sent in the clear. It keeps working for logged-out users, and the anonymous identifier persists for at least 27 days.
We flagged OpenAI's turn toward advertising months ago, when its plans first drifted from subscriptions and API fees toward selling ad inventory. This is what that looks like down in the wiring. And they never asked you: no notice, no docket, no form, just a pixel that switched on inside a product people talk to like a diary. On Hacker News the mood was less outrage than resignation: the company that dropped "open" from its own name is now doing the thing everyone feared, turning the assistant you talk to into an ad business that watches you. And the hashing defense only goes so far. Your email, phone and name get scrambled, sure, but your city and postal code go out in plain text.
Why this matters: The AI you chat with is now tracking what you read and buy on other sites, and it keeps doing it even when you're logged out. The label on the cookie says analytics; the behavior is a cross-site identity graph that scrapes data advertisers never chose to hand over. Companies love to say the customer owns its data. Here, nobody even got asked.
Action this week: Block third-party cookies and add a tracker blocker today, because the __obi cookie is set even when you're logged out, so an in-app ad opt-out is no guarantee it stops. When I have pressed vendors on "who owns my data," the answer has never once produced an actual export, and here it didn't even produce a disclosure. If you run a site that added OpenAI's pixel, audit what your pages expose in the DOM, since the scraping wasn't something you agreed to supply, and watch whether the FTC or a state attorney general treats an "analytics" label on an identity cookie as a deceptive practice.
daily.dev: ChatGPT now knows what you do on other websites, via an ad collector (September 2026)
buchodi.com: ChatGPT now knows what you do on other websites via an ad collector (September 2026)
ai-tldr.dev: OpenAI __obi ad tracker analysis (September 2026)
3. The AI your team codes with is heading for a two-trillion-dollar public offering.
Anthropic's revenue run rate went from about nine billion dollars to a hundred billion in a year, and it files to go public in November.
The next time your team renews its AI coding tools, you will be negotiating with a company on its way to one of the largest public offerings in history. Anthropic's annualized revenue run rate has gone from about $9 billion at the end of 2025 to roughly $65 billion by the end of July to a pace topping $100 billion this year, driven by businesses adopting Claude for coding and everyday work. It moved its planned debut from October to November so it can show third-quarter results to investors first, and bankers have told investors the raise could top $100 billion at a valuation near $2 trillion. Sam Altman, notably, says OpenAI won't go public in 2026, calling the timing ill-advised.
Yesterday we covered Anthropic's claim that Claude now leads about a quarter of its own research; today's number is that same story read off the balance sheet. That run-rate curve is really your competitors' AI spending, and now it's out where everyone can see it. The revenue is real, companies genuinely hand this much work to AI now. What's harder to swallow is the $2 trillion tag, which only pencils out if that growth keeps compounding at a rate almost nothing sustains, and the compute bills plus the peak-hour usage limits Anthropic has already had to impose on Claude say the strain is showing. An S-1 will finally force the disclosures a headline hides: how concentrated the customers are, what the gross margins actually look like, and probably a few numbers they'd rather not print.
Why this matters: The AI tool your team already codes with is heading for the biggest IPO in history, and the run-rate behind it's a bill headed for your budget. A jump from $9 billion to $100 billion in a year is the clearest signal yet of how much real work companies are handing to these tools, which means your per-seat costs are riding the same curve. Once a supplier is public and defending a $2 trillion valuation, "essential" stops being cheap.
Action this week: Get one answer in writing before your next renewal: what you pay per seat now, and what happens to that price once Anthropic is trading and has a margin to defend. When I have watched a vendor go from scrappy to indispensable, the price of indispensable always finds the customer eventually, so pull your team's actual usage, name the two or three tasks these tools now do end to end, and mark the one place a human still signs off, because that short list is your leverage when the invoice climbs.
pymnts.com: Anthropic targets November IPO as revenue surges (September 18, 2026)
axios.com: Anthropic's revenue on pace to top $100 billion (September 18, 2026)
bloomberg.com: Anthropic's annualized revenue to top $100 billion in 2026 (September 18, 2026)
PickBits Daily Signal is free. If it lands in your inbox every day and it's worth something to you, the best way to support it's to forward it to someone who would read it. Subscribe today!
4. An AI reading a cheap blood test can help predict who recovers after a brain bleed.
Researchers hit about ninety percent accuracy using D-dimer, a clot marker every hospital already measures.
Picture the family in an emergency room waiting to hear whether someone they love will walk out of a brain bleed. When an aneurysm ruptures and bleeds around the brain it strikes suddenly, often in midlife, and roughly one in four survivors never regains independent function, yet for decades that bedside call has leaned on clinical grading scales that leave a frustrating amount of uncertainty. A study in Neurocritical Care by Yanze Wu, Ping Hu and colleagues at Nanchang University shows a cheap, universally available input can sharpen it: D-dimer, a clot-breakdown marker your ER already measures in routine bloodwork. Training seven machine-learning models on 473 patients, the team found admission D-dimer carried independent prognostic weight, and an XGBoost model reached an AUC of 0.904 predicting poor recovery on held-out cases; a simpler composite of D-dimer, total bleeding volume and Hunt-Hess grade reached 0.867.
We've run a health-AI study almost every week this month, from AstraZeneca's Milton reading routine bloodwork to flag a thousand diseases years early, to models matching physicians on clean written-up cases. Skip the accuracy number for a second. The point here is the input: D-dimer is a test every hospital already runs, not a proprietary scanner, so a tool built on it could reach a rural or community ER that will never buy bespoke imaging AI. But it's early: single-center research on one hospital's patients, and prognostic models that shine on their own training cohort routinely stumble once they meet someone else's.
Why this matters: A cheap blood test you already get in the ER, read by AI, hit about ninety percent accuracy predicting who recovers after a brain bleed. The promise here isn't a smarter scanner, it's a smarter reading of a test every hospital already runs, which is exactly the kind of tool a clinic with no imaging AI could actually use.
Action this week: Ask the care team how they estimate recovery odds, by clinical grade, bleeding volume, and now maybe routine labs, rather than asking for an "AI score" that doesn't exist yet at the bedside. The milestone I am actually watching is external, multi-center validation, because a model this good on one hospital's 473 patients can still stumble on yours, and the whole promise falls apart if it's never tested on US and rural populations.
link.springer.com: D-dimer and machine learning for predicting functional recovery after aneurysmal subarachnoid hemorrhage, Neurocritical Care (September 2026)
pubmed.ncbi.nlm.nih.gov: D-dimer machine-learning prognosis after aneurysmal subarachnoid hemorrhage (2026)
» What to watch this week
- Whether a judge certifies the antitrust class, and whether the DOJ or FTC files a statement of interest. The case turns on whether "AI safety" coordination is a defense to a Sherman Act claim, and an antitrust exemption granted in Washington could moot the whole lawsuit.
- Whether OpenAI responds to the __obi findings and whether any regulator treats the cross-site cookie as a deceptive practice. Watch too for site owners auditing what their own pages leak into the pixel, since the scraping wasn't something advertisers deliberately supplied.
- Which disclosures Anthropic's S-1 carries when it files. Customer concentration, gross margins, and compute costs are the numbers the $100 billion run-rate headline hides, and they're what will tell you whether a $2 trillion tag is grounded.
- Whether the D-dimer prognosis model gets external, multi-center validation and is published openly. The entire promise is prognosis that reaches ERs that cannot buy proprietary imaging AI, and that only holds if the model is validated on US and rural populations first.
Tomorrow's signal lands here.