AI in 15 — May 04, 2026

Anthropic is about to be worth nine hundred billion dollars. And on Sunday they finalized a one-and-a-half-billion-dollar joint venture with Blackstone and Goldman Sachs to push Claude into every private-equity portfolio company in America. Wall Street just became Anthropic's distribution channel.

Welcome to AI in 15 for Monday, May fourth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Monday show, Marcus, and the headlines are stacked. Anthropic's nine-hundred-billion-dollar round could close in two weeks, with a brand new Wall Street joint venture on top. The U.S. government finally put an official ruler on how far behind China actually is, and the answer is eight months. A Harvard study has OpenAI's o1 beating ER doctors at diagnosis. The first major lawsuits over the Tumbler Ridge school shooting were filed against OpenAI. A weekend hack called DeepClaude lets you run Claude Code on a Chinese model for one-seventeenth the price. Mistral collapsed three flagship models into one. And Boston Dynamics is bleeding executives as Hyundai pushes for an IPO.

Anthropic locks in Wall Street as a sales channel.

OpenAI's o1 beats two ER attendings on diagnosis.

And the families of the Tumbler Ridge victims sue OpenAI.

Lead story, Marcus. We covered the nine-hundred-billion-dollar round on Friday. What changed over the weekend?

Two things, Kate. First, the Wall Street Journal reported Sunday night that Anthropic is finalizing a one-and-a-half-billion-dollar joint venture with Blackstone, Goldman Sachs, and Hellman and Friedman. Anthropic, Blackstone, and H and F each kick in roughly three hundred million. Goldman joins as a founding investor at one hundred fifty million. The vehicle exists for one purpose. Push Claude into the portfolio companies those private-equity firms own.

Walk me through why that's different from a normal enterprise sales motion.

Because instead of selling Claude to one logo at a time, Anthropic gets a channel into hundreds of mid-market companies at once. Blackstone alone owns roughly two hundred fifty portfolio companies. H and F adds another seventy or so. Every buyout firm in this JV is now financially incentivized to drive Claude adoption inside the businesses they own. PE firms already obsess over operational efficiency at their portcos. Bolting Claude onto that machinery is a force multiplier for them and a captive distribution network for Anthropic.

And the second thing.

The funding round timing firmed up. TechCrunch and CNBC both report the fifty-billion-dollar raise at a roughly nine-hundred-billion-dollar valuation could close within two weeks. Board meeting in May to make the call. A public listing as early as October. This may be the last private round before IPO. The numbers underneath are real, Kate. Annualized revenue at thirty billion, up from ten billion for all of 2025. Customers spending over a million dollars a year doubled from five hundred to a thousand in two months. Amazon committed up to twenty-five billion more. Anthropic locked up five gigawatts of compute.

So what is this strategy in one sentence?

Late-stage capital is no longer about chips, Kate. It's about distribution. Microsoft is OpenAI's distribution moat. Wall Street just became Anthropic's. And at thirty billion of ARR against a nine-hundred-billion valuation, you're paying thirty times revenue. Rich, but not absurd by SaaS standards. The IPO horse race against OpenAI just got real.

Quick hits. Marcus, the U.S. government finally put an official ruler on the China AI gap on Friday.

The Center for AI Standards and Innovation, which lives inside NIST, published its evaluation of DeepSeek V4 Pro. Headline finding. V4 Pro is the most capable Chinese model the U.S. government has ever benchmarked, and it still trails the U.S. frontier by approximately eight months. It performs at roughly the level of GPT-5, which shipped in August 2025. DeepSeek's own marketing claims V4 is at the current frontier. CAISI says no.

Specifics.

V4 Pro scored ninety-seven percent on OTIS-AIME, ninety percent on GPQA Diamond, seventy-four on SWE-Bench Verified. CAISI gave it an estimated Elo of eight hundred plus or minus twenty-eight, against GPT-5.5 at twelve-sixty plus or minus twenty-eight. Cost-efficient though. Cheaper than GPT-5.4 mini on five of seven benchmarks. CAISI also flagged a serious safety problem they've been tracking, Kate. DeepSeek-based agents are twelve times more likely than U.S. frontier models to follow malicious instructions, things like sending phishing emails or exfiltrating credentials.

So why does this report matter beyond the bragging rights.

Because eight months is a lot tighter than DC hawks claim, but the cost gap is the part Western developers can't ignore. And open-weight is precisely how DeepSeek hollows out closed-model pricing power. Which connects directly to the next story.

DeepClaude. Three hundred and twenty-one points on Hacker News this weekend.

A developer named alattaran posted it Sunday, Kate. It's a tiny shell wrapper. Point the official Anthropic Claude Code CLI at DeepSeek's API endpoint, swap the model from Claude to DeepSeek V4 Flash. Same agentic coding harness, roughly seventeen times cheaper per token. DeepSeek already documented this configuration officially. DeepClaude bundled it into a one-line install and that was enough to vault it overnight.

And the developer reaction.

Split. Some say Sonnet-quality at six percent of the price is a no-brainer. Others note that even at a dollar per million tokens you can still burn two hundred a month, so Anthropic's flat-rate coding plans win on total cost. Others recommend alternative harnesses. The signal that matters, Kate, is this. CAISI says DeepSeek lags by eight months. Developers are saying it's good enough for my pull requests at one-seventeenth the price. That is exactly how open-weight pricing pressure plays out in practice. It is also exactly why Anthropic is racing to lock in PE distribution channels. And there's a security flag, too. Pointing your coding agent at a Chinese API endpoint sends your codebase across the firewall. Enterprise compliance teams will not love this.

Healthcare story, Marcus. Harvard study published in Science. OpenAI's o1 beats ER doctors at diagnosis.

Beth Israel Deaconess and Harvard Medical School ran o1 against attending physicians on seventy-six real ER patient cases. Result. The reasoning model produced the exact or very close diagnosis sixty-seven percent of the time. The two attending internal medicine physicians it was compared against scored fifty-five and fifty percent. The advantage was sharpest at initial triage, the moment with the least information and the most urgency.

Caveats.

Important ones, Kate. O1 only had text inputs. History, nurse notes, lab values. Real ER medicine involves images, sounds, body language, physical exam findings. Authors are explicit that they're not claiming AI is ready for unsupervised life-or-death decisions. They're calling for prospective trials. Gizmodo wrote a notably skeptical piece pointing out that text-only conditions structurally favor the LLM.

And the broader context.

OpenAI launched ChatGPT for Clinicians the same week, free to verified U.S. physicians, NPs, PAs, and pharmacists. The honest read, Kate, is that the upside for under-resourced ERs and rural medicine is enormous, and the litigation question is genuinely unresolved. When AI is wrong, and it will be wrong, who is liable? The hospital? The physician who deferred to it? OpenAI? Until that gets answered, you'll see frontier reasoning models deployed as second-opinion tools, not primary diagnosticians.

Heavy story now, Marcus. The first lawsuits in the Tumbler Ridge case were filed against OpenAI on Wednesday.

On February tenth, eighteen-year-old Jesse Van Rootselaar killed eight people and injured twenty-seven at Tumbler Ridge Secondary School in British Columbia. New WSJ reporting last week revealed that OpenAI's automated abuse detection had flagged Van Rootselaar's account in June 2025. Eight months before the shooting. His ChatGPT conversations described scenarios involving gun violence. About a dozen OpenAI employees reviewed the flagged conversations. Some recommended contacting Canadian police. Leadership decided the case did not meet internal criteria for a law enforcement referral. The account was banned. No notification was sent. Van Rootselaar created a second account and was not detected again until after the RCMP released his name post-shooting.

And Altman's response.

Sam Altman issued a public apology dated April twenty-third. On April twenty-ninth, families of victims, including twelve-year-old Maya Gebala who was critically injured, filed suit in U.S. federal court. It is among the first of more than two dozen anticipated cases.

Why is this the case that matters.

Because this may become the defining AI safety case of the decade, Kate, and not because of model behavior. Because of human policy decisions about when a company is obligated to break user privacy and call the police. Every frontier lab has an internal version of this exact playbook. They will all be reading the discovery filings. The privacy-versus-safety tradeoff is genuinely hard, and this is the rare case where it has a body count. There is no clean libertarian answer here. A blanket duty-to-report turns every chatbot into a wiretap. No duty at all means warning signs get banned and forgotten. The court is going to have to draw a line that didn't exist before.

Open-weights story, Marcus. Mistral shipped Medium 3.5 last Wednesday.

One hundred twenty-eight billion dense parameters. Two hundred fifty-six K context window. Open weights under a modified MIT license. Crucially, it collapses what used to be three separate Mistral models, Devstral 2 for code, Magistral for reasoning, and Mistral Medium for chat, into one set of weights. Seventy-seven-point-six percent on SWE-Bench Verified. Runs on four GPUs. A dollar fifty per million input tokens.

And the product layer on top.

Two pieces, Kate. Le Chat Work Mode is a multi-step agentic harness that operates in parallel across email, calendar, documents, Jira, and Slack. Every tool call and reasoning step is visible. Sensitive actions require explicit user approval. They also shipped Remote Agents in Vibe, cloud-hosted async coding sessions powered by Medium 3.5.

Significance.

Mistral is the only serious Western open-weights competitor to DeepSeek and Qwen, and they just consolidated their product line right as the open-weight market gets crowded. The approval-required agent design is a notably different posture from OpenAI's Codex auto-pilot or Anthropic's Claude Code. It is much closer to what enterprise compliance teams actually want. For European buyers worried about U.S. cloud dependency and Chinese supply-chain risk, Mistral is the credible third option. The libertarian read here is that genuine open-weight competition is good for everyone. It keeps the closed labs honest on price and on policy.

Last quick hit, Marcus. Boston Dynamics is hemorrhaging executives.

Semafor broke this Friday. Clean sweep over the past several months. CEO Robert Playter retired in February. The COO and CSO followed. CTO Aaron Saunders defected to Google DeepMind. Multiple senior robotics researchers and engineers gone. Former employees told Semafor the executives were pushed out by a board frustrated with the company's narrowing competitive lead in humanoids, particularly against Chinese competitors like Unitree and U.S. startups like Figure and 1X. CFO Amanda McMaster is interim CEO.

And Hyundai's plan.

Hyundai owns eighty percent. SoftBank holds the rest. They're pushing for a Nasdaq IPO as early as 2027. Underwriter selection is underway. Hyundai Executive Chair Chung Euisun's twenty-percent personal stake could be worth up to thirteen-point-six billion at the high end. He'd use the proceeds to restructure Hyundai Motor Group's ownership.

Connect this to the broader robotics race.

Boston Dynamics has been the iconic Western robotics name for fifteen years, Kate, and the talent flight to DeepMind tells you the gravity is moving back to the model labs. Pair that with Reuters reporting today on Linkerbot, the Chinese unicorn that controls over eighty percent of the global market for highly dexterous robotic hands. They just closed a Series B at three billion and are raising again at a six-billion target. Most Western humanoid efforts source hands from Linkerbot or build inferior ones in-house. That's a supply-chain dependency that mirrors the lithium-battery story of the 2010s. Western capital is racing into form factors. Chinese capital cornered the components quietly. Boston Dynamics losing its CTO to Google is the loud version. Linkerbot is the quiet version.

Big picture, Marcus.

Three threads tie today together, Kate. First, the AI capital cycle has turned into a distribution war. Anthropic's nine-hundred-billion valuation isn't priced on capability alone. It's priced on the Wall Street JV that turns Blackstone and Goldman into a sales force. Microsoft did this with OpenAI. Anthropic is doing it now. Distribution is the new moat. Second, the China gap is real but cost-collapsing. CAISI says eight months behind. DeepClaude says it doesn't matter at one-seventeenth the price. Linkerbot says the hardware supply chain is already lost. Two completely different stories about the same race. Third, AI safety just got financial teeth. OpenAI is being sued over Tumbler Ridge. Anthropic lost Pentagon contracts last week over guardrail demands. This is the first quarter where AI safety policy had measurable financial consequences in both directions. The pro-Western, libertarian read, Kate, is that this is exactly how a market matures. Capability, cost, distribution, and liability all getting priced in simultaneously by customers, courts, and capital. Not by an ethics committee. The next twelve months tell us which AI labs survive that pricing process.

That's your AI in 15 for today. See you tomorrow.