AI in 15 — May 29, 2026

Sixty-five billion dollars raised in a single round. A nine hundred sixty-five billion dollar valuation. One Apple-sized leap away from a trillion. And Anthropic shipped it on the same day they confirmed they have a higher-tier model so capable they cannot release it yet because the cyber risk is too high.

Welcome to AI in 15 for Friday, May twenty-ninth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Big finish to the week, Marcus. Anthropic's Series H closes officially at sixty-five billion. Claude Opus 4.8 ships with a refreshingly modest pitch. Anthropic confirms a higher class model called Mythos is in limited distribution. Dynamic Workflows in Claude Code lands, and Jarred Sumner ported the entire Bun runtime from Zig to Rust in eleven days. Axios runs a sharp piece on AI sticker shock with Microsoft cancelling most of its Claude Code licenses. Altman publicly says he was, quote, pretty wrong about AI killing jobs. A San Francisco homeowner sues a humanoid robotics startup over a trashed Airbnb. A mystery Tencent model is hoovering up OpenRouter token volume. And frontier LLMs disagree on basic fact checks.

Anthropic prices in the IPO before it files.

Seven hundred fifty thousand lines of Rust in eleven days.

And the enterprise spending hangover gets a name.

Lead story, Marcus. The Series H is now official.

Confirmed yesterday, Kate. Anthropic raised sixty-five billion dollars at a nine hundred sixty-five billion dollar post-money valuation. Altimeter, Dragoneer, Greenoaks and Sequoia co-led. Capital Group, Coatue, D1, GIC and ICONIQ joined as co-leads. Fifteen billion of the package is previously committed hyperscaler money — Amazon alone wrote a five billion dollar cheque inside this round. Strategic partnerships with Micron, Samsung, and SK hynix on memory. Expanded compute capacity with Amazon, Google, Broadcom and SpaceX. And run-rate revenue crossed forty-seven billion dollars in early May, up from the thirty billion number we discussed yesterday.

Wait, forty-seven, not thirty.

Forty-seven, Kate. The thirty billion figure was the early-April annualized number that leaked through the Series H prep. The May disclosure puts it at forty-seven. That's an additional seventeen billion of annualized run-rate added in five weeks. Krishna Rao, the CFO, framed it as fuel to serve, quote, historic demand. The money is earmarked for safety and interpretability research, compute, and product scaling.

So the elephant in the room.

One Apple away from a trillion as a private company, Kate. Hacker News spent yesterday openly asking who actually gets paid back before the IPO, and whether private mega-rounds are pulling listings off the public exchanges entirely. The pro-Western libertarian read — these are real customers, real margins on the model side, and a revenue ramp very few public companies in history have matched. The uncomfortable read — Anthropic just priced itself ahead of its IPO at a number that requires forty-seven billion of run-rate to compound through enterprise pushback that, as you'll see, is now visibly arriving. They have effectively pre-sold the IPO to private capital. October retail money will inherit whatever discipline lands between now and then.

Quick hits. Marcus, Claude Opus 4.8.

Shipped the same day as the round, Kate. Pricing holds steady — five dollars per million input tokens, twenty-five per million output. A new Fast Mode at ten and fifty for two-and-a-half-times speed. The framing was the news. Anthropic described it as, quote, a modest but tangible improvement. That language is a real cultural shift. The substance — Opus 4.8 is roughly four times less likely than 4.7 to let code flaws pass unremarked. Gains in agentic reliability and tool efficiency. New highs on prosocial-traits alignment. Notably reduced misaligned behavior. The Messages API now accepts system entries mid-task without breaking prompt cache, and you can finally turn off adaptive thinking on the web UI. Simon Willison's pelican-on-a-bicycle benchmark shows 4.8 at high thinking visibly better than 4.7. Small but real.

And the buried headline.

Yes, Kate. Buried in the release notes — Project Glasswing graduates and the Claude Mythos Preview is now being distributed to a small number of organizations specifically for cybersecurity work. Anthropic explicitly says models at this capability level need stronger cyber safeguards before general release, expected, quote, in the coming weeks. So Anthropic is publicly acknowledging it has a model gated because of cyber-offense risk. That's a first. We've been tracking Mythos for weeks through the Apple CVE and the OpenBSD twenty-seven-year-old bug. Now they're saying it on their own news page. Two signals at once — frontier capability gains are getting tighter at the headline level, and the next class of model exists but is too dangerous to ship broadly today.

Dynamic Workflows, Marcus. The Bun port is the proof.

Genuinely impressive, Kate. Anthropic released Dynamic Workflows for Claude Code. The system lets Claude orchestrate tens to hundreds of parallel subagents inside one session. Claude writes the orchestration scripts itself, breaks the job into subtasks, fans work out, verifies independently, and iterates until convergence. Progress auto-saves. You either ask for a workflow explicitly, or you enable the new ultracode setting and Claude decides when to spin one up.

And the demo.

Jarred Sumner ported Bun — the entire JavaScript runtime — from Zig to Rust. Eleven days from first commit to merge. Roughly seven hundred fifty thousand lines of generated Rust. Ninety-nine-point-eight percent test pass rate against the existing test suite. The workflow assigned two reviewers per file, mapped Rust lifetimes for struct fields, wrote behavior-identical files, ran a fix loop until builds and tests passed, and used an overnight job to address data copies.

Caveats.

Two real ones, Kate. First — Bun had an extraordinarily high-coverage test suite to act as the forcing function. Most projects do not. If your repo has eight percent line coverage, this technique cannot tell you what it just broke. Second — and this connects directly to the next story — workflows consume, quote, substantially more tokens than normal Claude Code sessions. So this is the closest public demonstration yet of agentic AI doing serious systems-software rewrites, and it lands in the same news cycle as Microsoft cancelling Claude Code licences over cost. Both things are true.

AI sticker shock, Marcus. Axios.

Sharpest enterprise piece this week, Kate. Axios reports corporate leaders are starting to question whether soaring AI spending is delivering meaningful returns. The specifics — Microsoft cancelled most of its Claude Code licences, partly over cost. Uber's COO, who we covered Tuesday and Wednesday, said AI expenses are getting harder to justify. And an AI consultant told Axios about a single client that burned through five hundred million dollars in one month after failing to put usage limits on Claude licences. Staff were running unrestricted agentic workflows. Five hundred million. One client. One month.

Marcus, that's the head-exploding number for May.

It is, Kate. And the diagnosis Axios offers is structural. Four friction points — companies are automating things employees dislike rather than revenue drivers. Token costs run away on trivial queries. Leadership doesn't know how to implement. And over-restricted data access makes agents useless. GitHub Copilot reportedly ran a roadshow this week teaching enterprises to optimize token usage — token-maxxing in reverse. Amazon scrapped its internal Kiro-AI usage leaderboard after employees gamed it for badge points. A textbook Goodhart's Law moment. And one Hacker News commenter, very dry — quote — thanks to AI we're producing more code and more MRs, faster than ever, but the milestones aren't getting hit any sooner. Actually the opposite, if anything.

Stakes.

This is the counter-narrative to the forty-seven-billion run-rate, Kate. Anthropic's revenue may be coming from a smaller, more concentrated, more cost-conscious cohort than the valuation implies. The pro-Western libertarian read — markets discipline themselves. Microsoft pulling licences and consultants quoting five-hundred-million-dollar mistakes is exactly the feedback loop forming. The uncomfortable read — if the discipline arrives faster than the revenue compounds, the IPO window in October looks very different.

The walk-back, Marcus. Altman and Amodei.

Fortune piece this week, Kate. Receipts attached. June 2025, Sam Altman warned a lot of jobs would go away, especially entry-level white-collar. Dario Amodei said AI could eliminate roughly fifty percent of white-collar jobs. Both have visibly changed their tune. Altman now says he was, quote, pretty wrong. He tried delegating his own email and Slack to AI and found people don't actually want to interact with the bot. Amodei is reframing automation as a job multiplier — if you automate ninety percent of a role, the remaining ten percent, quote, expands to be one hundred percent. Ten-times productivity rather than ten-times layoffs.

Timing, Marcus.

Suspiciously convenient, Kate. Both labs are reportedly preparing 2026 IPOs at trillion-dollar valuations. Pew data referenced in the Hacker News thread — more than half of Americans are now more concerned about AI than excited about it. You don't IPO into that mood with a jobs-apocalypse pitch. So the rhetoric softens. The pro-Western libertarian read — this is also the first time these CEOs have publicly conceded the displacement story didn't unfold as predicted. The lived experience inside companies — slower milestones despite faster code generation — supports the walk-back. Both can be true. The framing has shifted because the data has shifted, and because the prospectus is being drafted.

Marcus, the Airbnb lawsuit. I cannot believe this is real.

It is very real, Kate. Sean Donovan, San Francisco homeowner, filed suit Tuesday in San Francisco Superior Court against The Bot Company — the two-billion-dollar humanoid robotics startup founded by Tesla and Cruise alumni. Donovan alleges Bot Company employees rented his four-bedroom home under false pretenses via Airbnb in April, then used the property as a household-chores test environment for prototype robots. His Ring camera caught people moving large black cases into the home on April twelfth. He came home eleven days later to a cracked refrigerator shelf, broken glass in the garbage disposal, stained linens, chipped bathroom tile, damaged furniture and scratched appliances. He's seeking twelve thousand three hundred eighty-three dollars and fifty cents plus lost commercial-use revenue.

First lawsuit of its kind.

As far as I can tell, Kate. First public allegation that a major embodied-AI company is doing covert in-home prototype testing in residential rentals. Real consent and disclosure questions. Hacker News commenters noted Bot Company likely doesn't even carry commercial insurance for these rentals. Move fast and break things, but in someone else's kitchen. As humanoid startups race toward household-chore demos this gets worse before it gets better.

Marcus, the mystery of Hy3.

Real puzzle, Kate. Tencent's Hy3 Preview — open-source, two hundred ninety-five billion parameter mixture-of-experts model with twenty-one billion active — has rocketed to the top of OpenRouter's volume rankings. Between April twenty-third and May twelfth it consumed seven-point-seven trillion tokens. More than Kimi K2.6 at five-point-zero, more than Claude Sonnet 4.6 at four-point-zero, more than Claude Opus 4.7 at three-point-four. And yet — almost nobody is talking about it. Almost no Reddit or Hacker News discussion. Its own published benchmarks show it underperforming other Chinese open-source models on coding. Simon Willison got weird HTML quirks. Others report it gets stuck in thought loops and underperforms a thirty-one-billion-parameter local model.

So what's going on.

Max Woolf's best guess, Kate — a single large non-agentic application is using Hy3 as a cheap data-processing backbone. Once those users discover DeepSeek's effectively-lower rate, they'll switch. And the critical context Simon Willison added — OpenRouter only sees tokens routed through OpenRouter. Most Anthropic API users go direct. The leaderboard is not a global usage measure. Worth flagging — a Chinese frontier-tier MoE model is racking up massive top-of-leaderboard volume via essentially-free pricing. Whether that's organic demand or a strategic loss-leader to pull mindshare and data away from US labs is the open question. It fits the pattern. Tencent gets to claim leaderboard dominance in its Q1 results presentation regardless of the actual quality story.

Last quick hit, Marcus. Frontier LLMs disagree on fact-checks.

Real-world experiment from Lenz dot io, Kate. They ran user-submitted fact-check claims through GPT-5.4, Opus 4.7, Gemini 3, Gemini 3 with retrieval, and Sonar Pro. Forced four-way classification — True, Mostly True, Misleading, False — no explanation. Disagreement was significant. And not just on contested political claims. On plain factual ones. The viral example — extraterrestrial life exists somewhere in the universe. GPT-5.4 and Opus 4.7 said Misleading. Gemini 3 and Sonar Pro said False. Methodology has real problems — no Unknown option, forced classification — but the headline lands. Frontier models disagree on the same basic claims, and the disagreement is not predictable from model size or vendor.

Why it matters.

As LLMs get embedded in news, search and content moderation pipelines, Kate, which-model-you-ask becomes an epistemic question. Build a fact-checking layer on Gemini and your competitor builds on GPT, you'll classify the same claim differently — and neither layer tells you it's uncertain. That's an under-discussed risk for every AI-powered content product shipping right now.

Big picture, Marcus.

One coherent story to close the week, Kate. Anthropic raises sixty-five billion at near-trillion-dollar valuation while shipping an explicitly modest model update. Microsoft cancels Claude Code licences over cost. Axios reports CFOs questioning ROI. Altman and Amodei walk back the same jobs-apocalypse rhetoric they used to attract capital. Meanwhile Dynamic Workflows shows what the technology can really do — Bun ported to Rust in eleven days, seven hundred fifty thousand lines, ninety-nine-point-eight percent test pass — and Mythos sits gated for cyber risk in a small partner program. Under it all, a free Tencent MoE hoovers token volume and a humanoid robot startup allegedly trashes an Airbnb. The contradiction is the story. Capability is rising. Valuations are rising. Enterprise discipline of, is this actually paying off, is rising at the same time. The pro-Western libertarian read — that's a functioning market. Buyers push back. Sellers adjust. The labs that ship real revenue keep raising. The ones that don't will get re-priced. The uncomfortable read — the discipline has to arrive before October, or it arrives on retail. Next week we'll watch for whether any other Fortune fifty echoes Microsoft on Claude Code, whether Mythos enters wider distribution on Anthropic's coming-weeks timeline, and whether the Bot Company files an answer that explains what those large black cases actually were.

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