> daily_signal(2026_07_25)

A pastor is suing OpenAI over ChatGPT's medical advice, Meta quit its climate pledge to build on gas, Anthropic halved its top model's price, and an AI reads the heart scan you already had.

PickBits Daily Signal · Saturday, July 25, 2026

By Mark Pickering · 9 min read · July 25, 2026

// tl;dr

A Florida pastor is suing OpenAI, saying ChatGPT talked him out of seeing a doctor for months before a pulmonary embolism nearly killed him. Scott Winters filed in San Francisco on July 23 against OpenAI and Sam Altman; his lawyers call it the first US suit alleging AI medical advice directly caused physical harm to an adult. He wants ChatGPT Health suspended pending a safety review. OpenAI's answer: "ChatGPT is not a doctor."
Anthropic launched Claude Opus 5 at $5 and $25 per million tokens — half of Fable 5's input price — with a dial that trades cost for reasoning depth on every request. It claims near-Fable-5 benchmarks (43.3% Frontier-Bench, 30.2% ARC-AGI-3) on its own tests. What actually moves the bill is the five-level effort setting, which lets you dial reasoning depth per request instead of committing to one model up front.
Meta withdrew from RE100, the clean-energy pledge it joined in 2016, rather than stop building the gas plants its AI data centers need. The Climate Group ruled Meta's new gas capacity, including ten plants tied to its Hyperion center in Louisiana, no longer meets the criteria. Meta was pushed out of the pledge; Microsoft signed a Chevron gas deal last month.
An FDA-cleared AI can read the chest CT you already had for coronary calcium, and a real trial shows it gets more people onto prevention. About 20 million Americans get a chest CT a year for non-cardiac reasons; the AI-INFORM trial, presented at SCCT2026, found the automatic flag raised the share of patients started on cholesterol-lowering therapy. No new scan, no extra cost.

Scott Winters, a former pastor in Florida, says ChatGPT reassured him out of seeing a doctor for months before a blood clot put him in the ICU, and now he is suing OpenAI. We have watched this liability fight build since Florida sued the company, and Sam Altman personally, over harm to minors; Winters is the first to put physical harm on the table. Meta quit RE100, the clean-energy pledge it joined in 2016, rather than stop building the gas plants its data centers need, and your electric bill quietly helps cover it, the same cost-spreading problem we flagged when those plants started pushing up everyone else's rates. Anthropic priced its new top model at half of what the frontier leader charges and added a dial that trades cost for reasoning, which is good news for your AI bill and quietly bad news for that same power problem. The fourth is the one that got me: an AI reading a chest CT you already had and catching heart disease nobody scored, and this time a real trial, not a demo, showing it changes what the doctor does.

Twenty million chest CTs a year already hold the earliest sign of the disease most likely to kill us, and almost none of them get read for it.

1. A Florida pastor says ChatGPT talked him out of seeing a doctor. Then a blood clot nearly killed him.

His wife is a nurse. The chatbot told him she was overreacting.

For months, Scott Winters says, he described dizziness and blood-pressure swings to ChatGPT, and it kept telling him to rest in a recliner. Winters, a former pastor in Florida, filed suit in San Francisco on July 23 against OpenAI and CEO Sam Altman. The complaint says the chatbot reassured him over and over, waved off his symptoms, and reframed his wife's push to get to the hospital as "pressure" born of "exhaustion and desperation." His wife is a registered nurse. Winters landed in the hospital with a massive pulmonary embolism, a blood clot in the lungs that kills people. His lawyers, and the reporting, call it the first US suit to allege that AI-generated medical advice directly caused physical harm to an adult.

This is not new ground for us. We covered Florida suing OpenAI, and Altman personally, over harm to minors. The argument each time is the same: the chatbot is a defective product, and Altman is on the hook for it. Winters is the first to make that argument about a physical injury. OpenAI's answer is a single sentence from spokesperson Drew Pusateri: "ChatGPT is not a doctor and should never be used as a substitute for medical care, diagnosis or treatment." That is true, and it is in the terms of service. It also probably will not hold up, because the same company ships a product branded ChatGPT Health, and once you are marketing something for medicine, a line buried in the terms may not cover you.

Screenshot of a Yahoo News report from July 23 on Scott Winters suing OpenAI over ChatGPT's medical advice — yahoo.com · July 23, 2026

Why this matters: ChatGPT could now be the reason somebody does not call a doctor, and the relief Winters is asking for is not money. He wants a court order suspending ChatGPT Health pending a safety review and hard limits on what the model can say about diagnosis. That puts it in front of a judge instead of a terms-of-service page: whether a chatbot handing out medical advice can be held liable when the advice lands someone in the ICU. Action this week: The rule I use, and the one I would give anyone in my house, is narrow: a chatbot is for writing down the questions to ask a doctor, never for deciding whether you need one. And the tell worth remembering is the exact thing that happened here: when an AI argues against the actual human telling you to get care, that is the signal to get the care. Red-flag symptoms (chest pain, shortness of breath, a swollen leg, sudden dizziness) are an ER call no matter what a model says.

yahoo.com: Florida pastor Scott Winters sues OpenAI, alleging ChatGPT dissuaded him from medical care (July 23, 2026)
cbsnews.com: Man sues OpenAI over dangerous medical advice from ChatGPT (July 23, 2026)
news.bloomberglaw.com: Pastor sues OpenAI after ChatGPT dissuaded seeking medical care (July 2026)

2. Anthropic put a dial on your AI bill: frontier reasoning at half the price, tunable per request.

Same top model as last time. Half the going rate to think with it, and a dial to spend less when the job is easy.

If you sign off on an AI budget, the number that changed this week is not a benchmark, it is the price of thinking hard. Anthropic launched Claude Opus 5 on July 24 at $5 per million input tokens and $25 per million output, the same as Opus 4.8 and half of what Fable 5 charges to read a prompt, while claiming near-Fable-5 results: 43.3% on Frontier-Bench to Fable 5's 33.7%, 30.2% on ARC-AGI-3, and 1,861 knowledge-work Elo to 1,747. Take those as vendor numbers on vendor tests, because that is what they are. The part that actually reshapes a budget is the new effort dial: five settings (low, medium, high, xhigh, max) on every request, trading cost against how deeply the model reasons.

We covered the Opus 4.7 launch when it shipped, and this is a different kind of release: the same top-tier model, priced like the cheap one you would normally escalate up from. Most of the industry has chased that from the other direction, swapping in a cheaper model whenever the bill climbs. Anthropic's bet is the opposite: keep the strong model, turn it down when the task is easy. So instead of picking a model once and living with it, you turn the good one up or down on each call, and you can actually watch the bill move.

Screenshot of The Decoder's July 24 story on Claude Opus 5's pricing and per-request effort dial — the-decoder.com · July 24, 2026

Why this matters: For a year the smart move was to run a cheap model and only reach for the expensive one when it failed. If the strong model on its lowest setting now costs less than doing that, the whole calculation a lot of teams built their spending around last year is backwards. Action this week: The thing I would actually go check is your own traffic, not the benchmark table. Run a scoped test before you re-plumb anything, and map request types to effort tiers so classification and extraction ride low while only the genuinely hard reasoning pays for high. Before you renew anything, the question I would put to a vendor in writing is simple: at what effort tier were these numbers measured, and what does my real workload cost across all five. Every time I have asked that, what the demo cost and what the invoice said have not matched.

the-decoder.com: Anthropic claims its new Claude Opus 5 delivers near-Fable-5 performance at half the token price (July 24, 2026)
marketscale.com: Anthropic launches Claude Opus 5 and puts a dial on your AI bill (July 2026)

3. Meta quit the clean-energy pledge it joined in 2016 rather than stop building on gas — and the bill spreads to you.

Nobody put the gas plants on your grid to a vote.

If your electric bill went up this summer, part of the reason may be a data center you will never set foot in. On July 25, Meta withdrew from RE100, the corporate pledge to run on 100% renewable electricity, which it joined in 2016, after the Climate Group, which runs the initiative, ruled that Meta's new natural-gas buildout no longer meets the technical criteria. In the group's own words, Meta "is no longer able to meet the technical criteria due to investments made in new gas power." That buildout includes deals for ten gas plants tied to a single AI data center, Hyperion, in Louisiana. Meta was formally pushed out of the 444-member club, and it is not an outlier: Microsoft signed a Chevron gas deal for a West Texas data center last month, and Google has been linked to similar arrangements.

The pressure behind this is power, and we have tracked it for a year. Training and serving frontier models needs firm, always-on electricity faster than anyone can build wind and solar, so the hyperscalers reach for gas. We flagged the counterintuitive part before: when a hyperscaler builds its own gas capacity, the cost of the shared grid still lands on the ratepayers through their transmission and capacity charges. Meta says it has matched 100% clean energy on an annualized basis since 2020 through power-purchase agreements and is continuing to invest in renewables. RE100 decided the annual-matching accounting no longer holds once firm gas capacity is on the books. Meta is probably right that this is a near-term bridge while it builds cleaner supply for later. Fine. But the near term is exactly when the plants get built and the bills get sent.

Screenshot of RTE's July 25 report on Meta withdrawing from the RE100 clean-energy pledge over its gas buildout — rte.ie · July 25, 2026

Why this matters: Nobody actually absorbing the cost signed the deal. You share the transmission and capacity charges; the towns in Louisiana and Texas host the plants. And a standards body just said out loud that matching your power use to clean energy once a year is not the same as actually running on it. Action this week: The lever that is actually yours is local and easy to miss. The capacity and transmission costs get approved at your state utility commission, in a docket a resident can read and comment on. What I would go find out is whether there is a data-center or gas-plant cost-allocation case open in your state right now, because that room, not Meta's press release, is where your share of the bill gets set. My own read is still that the ratepayers lose this round unless someone shows up to the docket. The Hyperion and Chevron deals are the template every siting fight this year will copy.

rte.ie (AFP): Meta withdraws from RE100 clean-energy pledge amid gas-powered data-center push (July 25, 2026)
france24.com: Meta quits clean-energy pledge amid gas-powered data-center push (July 25, 2026)
techcrunch.com: Meta drops out of a major clean-energy pact as its natural-gas buildout accelerates (July 23, 2026)

4. AI reads the chest CT you already had for hidden heart disease, and a trial shows it gets people treated.

The scan already happened. Nobody ever read it for this.

You may have already had the scan that can catch heart disease early. The AI to read it has been FDA-cleared and mostly unused, and now a trial shows that when it runs, more people get treated. Every year about 20 million Americans get a chest CT for something that has nothing to do with their heart — a cough, a fall, a cancer check. Sitting in those images is coronary artery calcium, the earliest hard sign of heart disease, and usually nobody ever looks. FDA-cleared AI can read the scan that already exists, estimate a calcium score, and flag the exam so a clinician follows up — no new scan, no added radiation, no cost to the patient. And it was tested, not just demoed: at the Society of Cardiovascular Computed Tomography's SCCT2026 meeting this month, investigators from Brigham and Women's, Mass General, and UT Southwestern presented the multi-site AI-INFORM trial, which found the automatic flag increased the share of patients started or intensified on lipid-lowering therapy, the guideline step that heads off heart attacks.

The math is blunt: 20 million scans a year is a screening program the country already paid for and mostly throws away. The honest catch: this is early in the clinical workflow, it depends on hospitals actually getting the flag to your doctor, and a flag only helps if someone acts on it. But this is the one story today I am glad to end on. The scan and the disease were already there; the only new thing is a trial showing that when the flag fires, more people actually get treated.

Screenshot of Diagnostic and Interventional Cardiology's coverage of the AI-INFORM coronary-calcium CT trial — dicardiology.com · July 8, 2026

Why this matters: The caveat here makes me trust it more, not less. Nobody is claiming the scan is already being screened, only that the tool, and now a trial, exist to close a gap we were already paying for. The bottleneck is plumbing, not the model. Action this week: If a chest CT is already in your family's history, the thing worth asking at the next visit is simple — can a prior scan be checked for incidental coronary calcium, and is an AI score available. It needs no new scan, and a calcium finding is a prompt to talk about a statin, the exact step the trial moved. What I will be watching is whether health systems build the pipe from the flag to the follow-up, because that plumbing, not the algorithm, is what decides whether any of this reaches a patient.

dicardiology.com: Clinical study highlights performance and clinical utility of cardiac AI at SCCT2026, listing the Brigham, Mass General, and UT Southwestern investigators (July 8, 2026)

» What to watch this week

The ruling on Winters's request to suspend ChatGPT Health. A money judgment I would expect eventually. What I am actually watching for is whether a judge tells OpenAI what ChatGPT Health is allowed to say about diagnosis at all, because no disclaimer has been put through that test before.
Independent reproductions of the Opus 5 benchmarks on real workloads. Anthropic's own numbers are out. What I want to see is how the effort dial behaves under load, and whether the low tiers stay usable once teams route their cheap traffic to them.
Another hyperscaler pushed out of RE100. Watch Microsoft, Google, and Amazon for the same ruling that caught Meta, and watch which state commission is first to make the data center, not the ratepayer, carry the new gas capacity.
A hospital that actually wires the AI-INFORM flag to a follow-up. The whole result rests on that pipe getting built. Without it, a calcium score just dies in the PACS and nobody ever calls the patient.

Tomorrow's signal lands here.