The Meltdown — April 14–19, 2026
Six stories. Two of them are about the same company, and that company is having a strange week.
01 — Anthropic's Unreleased Model Put Treasury Secretary Bessent in a Room With Wall Street
Anthropic's Mythos model — the one it won't release publicly because it's too dangerous — reached Washington in a big way this week. Bloomberg reported on April 14 that the US Treasury's tech team was seeking access to Mythos to hunt for vulnerabilities in financial infrastructure. Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell separately convened a group of Wall Street executives to discuss the cybersecurity risks. Jack Clark, Anthropic's co-founder, confirmed the company briefed the Trump administration directly. What makes this strange: CISA, the federal cybersecurity agency best positioned to respond, has been gutted by the current administration. The government is seeking help from a private AI lab to defend against a threat that same lab created, while dismantling the agency it would normally call.
02 — Anthropic Is Worth $800 Billion, and It's Turning the Money Away
On April 14, Bloomberg reported that Anthropic has received investor offers valuing it at roughly $800 billion, more than double the $350 billion valuation it carried into February's $30 billion raise. The company has so far declined. The offers reflect extraordinary revenue growth: Anthropic crossed $30 billion in annual run-rate revenue in early April, representing roughly 1,400% year-over-year growth. The most likely explanation for the refusal is that an IPO makes more sense at this scale than another private round. The less obvious read is that Anthropic may be avoiding the obligations that come with a new set of investors right as it manages the Mythos situation and government scrutiny. Both could be true at once.
03 — The Anti-AI Backlash Moved From Polls to Molotov Cocktails
A Washington Post survey found Virginians comfortable with data centers in their community at half the level they were three years ago, and a Quinnipiac poll put national opposition at 65%. This week the situation escalated badly when a man threw a Molotov cocktail at the gate of Sam Altman's San Francisco home, and an Indianapolis City-County Council member's house was shot 13 times with a note reading "No Data Centers" left at the scene. Daniel Moreno-Gama, 20, of Spring, Texas was arrested for the Altman incident. The backlash has been building for two years, driven by energy costs, water usage, job anxiety, and tax incentives that mostly benefit corporations rather than the communities hosting the infrastructure, and now it's crossed into violence. Altman and his peers are about to discover that the people building the future have to live somewhere too.
04 — Stanford: The US Lead Over China in AI Is Now 2.7%
Stanford HAI released its 2026 AI Index this week, and the headline number is really stark. The US leads China in AI benchmark performance by 2.7%, down from a gap of 17 to 31 percentage points in 2023. US and Chinese models have traded the top spot multiple times since early 2025, with Anthropic's Claude Opus 4.6 currently leading, followed closely by ByteDance's Dola-Seed 2.0. A 2.7% margin flips on a single model release from either side. The report also noted that H-1B researcher inflows to the US dropped 89% over the past seven years, with 80% of that decline happening in the past year alone, partly attributed to new $100,000-per-hire employer fees. The US is simultaneously leading by less and making it harder for the people who could close the gap to stay.
05 — Anthropic Made Claude Code a Scheduling Platform, and Then Quietly Turned Down Its Own Model.
On April 14, Anthropic shipped a complete redesign of the Claude Code desktop app and launched "Routines" in research preview: cloud-hosted automations that run on a schedule without the user's machine being online, with access to repos and API connectors. The direction seems clear — Claude Code is becoming a background automation layer, not just a coding assistant, and it's a direct challenge to tools like GitHub Actions. The timing is awkward, though. The same week, Anthropic faced a serious backlash over quietly reducing Claude's default effort level to save on token costs. Stella Laurenzo, a Senior Director at AMD's AI group, filed a detailed GitHub complaint backed by analysis of nearly 7,000 Claude Code sessions documenting the regression in complex engineering tasks. Anthropic is selling more capability with one hand while dialing back performance with the other, and developers noticed.
06 — A Startup Taught a Robot to Fold Laundry It Had Never Seen Before
Physical Intelligence released π0.7 on April 16, and the claim is specific enough to take seriously: a robotic foundation model that demonstrates compositional generalization, combining skills from different training domains to solve problems it was never directly taught. The company tested it against the specialist models it had previously built for individual tasks and found the generalist matched their performance across coffee-making, laundry folding, and box assembly. The model accepts language coaching, visual subgoals, and strategy metadata at inference time to steer it through unfamiliar situations. Physical Intelligence is reportedly valued at around $11 billion. The larger implication is that the "one model, all tasks" approach in robotics might actually work, which would change the economics of the field considerably — making a fleet of general-purpose robots cheaper to train than a library of specialists.