A dense day in artificial intelligence: new model revelations, a massive infrastructure arms race, military AI integrations, and fresh signals about where the industry's power dynamics are shifting.
GPT-5.5 Arrives With Benchmarks and Caveats
OpenAI's GPT-5.5 completed a full multi-step cyberattack simulation end-to-end, achieving roughly 71% success on expert-level tasks — a result that prompted the UK's AI Security Institute to declare it competitive with, or better than, Anthropic's unreleased Mythos on advanced cybersecurity benchmarks. The NSA is now independently testing Mythos for hunting vulnerabilities in Microsoft software. The agency was reportedly impressed by its raw speed.
Despite its benchmark performance, GPT-5.5 comes with a significant asterisk: it ranks lower than competitors in subjective head-to-head comparisons and has shown elevated rates of confident hallucination — a pattern where the model delivers incorrect output without flagging uncertainty.
Anthropic's Jupiter Surfaces Early
A new Anthropic model labeled *claude-jupiter-v1-p* appeared in red-team testing ahead of the company's "Code with Claude" developer conference scheduled for May 6. The early leak suggests Anthropic is preparing a significant release, though details about Jupiter's capabilities remain scarce.
The Infrastructure Arms Race
Big Tech's Q1 capital expenditure numbers arrived, and they are staggering. Amazon, Microsoft, Meta, and Alphabet collectively spent more than $130 billion on AI infrastructure in a single quarter, with projections of up to $725 billion combined for the full year of 2026. OpenAI separately announced it has reached its 10-gigawatt compute target ahead of schedule.
SoftBank plans a $100 billion IPO for a new AI and robotics subsidiary called Roze AI, which will deploy autonomous robots to build data centers more efficiently. Meanwhile, analysts note that falling behind on AI infrastructure is increasingly treated as an extinction-level threat — a psychology that compels continued escalation regardless of whether current AI revenue justifies the spend.
Military AI Expands Its Footprint
Google amended a $200 million contract with the Pentagon, allowing it to run Gemini models on classified networks for "any lawful government purpose" — a move made over internal ethics objections. The platform GenAI.mil, which already has 1.3 million DoD personnel users, has generated tens of millions of prompts and deployed hundreds of thousands of AI agents in its first five months.
Chips, Markets, and Geopolitics
Taiwan surpassed Canada as the world's sixth-largest stock market, driven largely by TSMC, which reached a $1.8 trillion valuation on AI-driven semiconductor demand. Samsung reported a 750% increase in operating profit year-over-year, attributing the surge to AI-powered memory demand.
Huawei expects its AI chip sales to reach $12 billion in 2026 — 60% growth — and is on pace to become one of the top three AI chip suppliers globally. The White House simultaneously warned that China-linked actors are attempting to steal American AI through industrial-scale distillation using proxy accounts and millions of automated model interactions.
Models, Agents, and Product Launches
AWS will resell OpenAI models — including GPT-5.4 and Codex — through Amazon Bedrock after Microsoft lost its exclusive licensing arrangement.
Musk admitted in court that xAI partly trained Grok using distillation from OpenAI model outputs, confirming a widely suspected industry practice.
Kimi K2.6, from Moonshot AI, launched with 1 trillion parameters, longer autonomous coding sessions, and support for up to 300 parallel subagents.
Manus launched Cloud Computer, a persistent Ubuntu-based environment that stays online 24/7, letting AI agents retain files and execute background processes without interruption.
1X began full-scale production of its NEO humanoid robot at a 58,000 sq ft Hayward, California facility, targeting 10,000 units this year and 100,000 by end of 2027.
Healthcare, Biology, and Science
Google DeepMind unveiled an AI "Co-Clinician" designed to support doctors in diagnosis and evidence synthesis. In testing across 98 primary care scenarios, it produced zero critical errors in 97 of them. Separately, Zuckerberg-backed CZ Biohub committed $500 million to build better AI simulations of the human body, with the long-term aim of curing human disease.
Novo Nordisk signed a company-wide AI deal with OpenAI covering drug discovery, clinical trials, manufacturing, and supply chain operations.
Security and Legal Flashpoints
A critical Linux kernel vulnerability was disclosed: a logic bug exploitable with a 732-byte Python script can obtain root access on nearly all Linux distributions shipped since 2017. The discovery was AI-assisted. In the crypto space, April 2026 saw $629.69 million stolen from protocols in a single month — the highest on record — and a16z research demonstrated that AI coding agents can recreate exploits 70% of the time when given structured attack pattern knowledge.
A federal court ruling established that AI-generated advertising content can expose platforms to securities fraud liability when the AI exerts "ultimate authority" over what is said.
The Bigger Picture
China deployed 8,500 AI-powered robots across 26 provinces to modernize its power grid, backed by a $1 billion State Grid initiative. Meta announced plans to spend up to $145 billion in 2026, despite investor unease. Google Gemini began rolling out to millions of 2022+ GM vehicles. And approximately 39% of the roughly 11,000 new podcast feeds created in the nine days prior to May 1 may have been AI-generated — a phenomenon critics are calling "podslop."
Silicon Valley, meanwhile, is grappling openly with the implications. Many insiders believe advanced AI could significantly erode workers' economic power, with some predicting the emergence of a "permanent underclass" as routine cognitive work is automated away.