This Week in AI: April 6-12, 2026 (Week 15)

News of the week: April 6-12, 2026 (Week 15)

The week was dominated by Anthropic’s Claude Mythos Preview, a model that can find security flaws in software on its own, which is why the company has made it available only to a small circle of partners. Its capabilities prompted an emergency meeting between the heads of major US banks, the Treasury Secretary and the Fed Chair. Meta unveiled Muse Spark, the first model from its superintelligence lab, and OpenAI added a cheaper ChatGPT Pro plan. The race for compute continued: Anthropic signed massive deals with Broadcom, Google and CoreWeave, and Intel joined Elon Musk’s Terafab project. On the regulatory front, xAI sued over a Colorado law and Florida opened an investigation into OpenAI.

New Models

Claude Mythos Preview: a model too powerful for the public

On April 7, Anthropic introduced Claude Mythos Preview, a model that will not reach ordinary users for now. The company has made it available only to partners in its Project Glasswing initiative, who are meant to use it to find and fix vulnerabilities in their own systems before attackers discover them (PCWorld).

The partners include AWS, Apple, Google, Microsoft, Nvidia, CrowdStrike and the Linux Foundation, among others. According to Anthropic, during testing the model found thousands of so-called zero-day vulnerabilities, meaning flaws the vendor did not yet know about, across every major operating system and browser (Security Affairs).

Among the disclosed findings is a roughly 27-year-old bug in the networking code of OpenBSD.
Another is a 16-year-old bug in the H.264 codec of the FFmpeg library, which automated testing tools (fuzzers) had missed even though they had passed through that part of the code millions of times (VentureBeat).
The vast majority of the bugs found are still under coordinated disclosure, meaning they are awaiting a fix before they can be discussed publicly.

Meta Muse Spark

On April 8, Meta released Muse Spark, the first model from its new Meta Superintelligence Labs, into which Mark Zuckerberg poured billions of dollars last year. The model powers the Meta AI assistant in the app and on meta.ai, and is set to come to WhatsApp, Instagram, Facebook and Meta’s glasses in the coming weeks (Fortune).

Unlike the earlier Llama models, which developers could download for free, Muse Spark is closed. According to Business Standard, it shared fourth place in the overall Artificial Analysis ranking, making it a competitive but not leading model (Business Standard).

Zhipu GLM-5.1 as an open model

On April 8, the Chinese company Zhipu (brand Z.ai) released the weights of its flagship model GLM-5.1 under the MIT license, which also permits commercial use. The company claims it outperformed Claude Opus 4.6 on some coding benchmarks; independent verification is still lacking (CnTechPost).

Interestingly, Zhipu at the same time raised the price of access to its models through its API by about 10%. Chinese players are thus gradually moving away from a lowest-price strategy.

Products and Agents

Claude Managed Agents and general availability of Cowork

On April 8, Anthropic launched a public beta of Claude Managed Agents. It is a set of interfaces that lets companies build and run AI agents in the cloud, meaning programs that carry out multi-step tasks on their own. The service offers sandboxed code execution, persistent state storage, credential management and scoped permissions (InfoWorld).

A day later, on April 9, Claude Cowork, a tool that works with files and applications on the user’s computer, came out of its testing phase. It is available on all paid plans on both macOS and Windows and brings permission management for enterprise administrators (eWeek).

ChatGPT Pro for $100

On April 9, OpenAI added a new ChatGPT Pro plan for $100 a month, positioned between the $20 Plus plan and the existing $200 Pro plan. It offers five times the usage of the Codex coding tool compared with Plus, and through the end of May even ten times as much (MacRumors).

Commentators see the plan as a direct response to Claude Max at the same price, which is popular among developers and so-called “vibe coders,” people who program mainly by giving instructions in plain language.

Business and Investment

Anthropic: revenue above $30 billion and gigawatts of chips

On April 6, Anthropic announced that its annualized revenue (run rate, meaning current revenue extrapolated over a full year) had surpassed $30 billion, up from roughly $9 billion at the end of 2025. At the same time, it signed a deal for about 3.5 gigawatts of compute capacity on Google TPU chips, to be supplied by Broadcom starting in 2027 (The Register).

According to Broadcom’s filing, the deal runs through 2031 and most of the new capacity will be built in the US (Tom’s Hardware). The company also said it has more than a thousand business customers paying over $1 million a year (ITPro).

CoreWeave signed two major contracts in 48 hours

On April 9, cloud GPU provider CoreWeave expanded its contract with Meta by roughly another $21 billion, running through December 2032 (Bloomberg). Right after, on April 10, it announced a multi-year agreement with Anthropic for Nvidia chip capacity in US data centers; financial terms were not disclosed (CoreWeave).

Intel joins Musk’s Terafab project

On April 7, Intel announced that it is joining the Terafab project championed by Elon Musk’s companies SpaceX, xAI and Tesla. The goal is giant chip factories for artificial intelligence, robotics and orbital data centers; Intel is to contribute chip design, manufacturing and packaging (Tom’s Hardware).

The announcement still lacks details on how and when production is supposed to start. Intel shares nonetheless rose sharply on the news (Engadget).

Amazon plans to spend $200 billion this year

In his annual letter to shareholders on April 9, Amazon CEO Andy Jassy defended planned capital expenditures of around $200 billion for 2026, mostly on AI infrastructure. He also revealed that AI revenue in the AWS cloud had reached an annualized $15 billion (Constellation Research).

According to the letter, Amazon’s in-house Graviton, Trainium and Nitro chips generate more than $20 billion a year, and Amazon does not rule out selling them to other companies.
Amazon shares rose more than 5% after the letter was published (Modern Retail).

Regulation and Law

OpenAI proposes a “new social contract”

On April 6, OpenAI published a thirteen-page document, Industrial Policy for the Intelligence Age, with more than two dozen proposals for the era of superintelligence. They include a public wealth fund, a shorter workweek and higher taxes on capital gains; the company describes them as “intentionally early and exploratory” (The Decoder).

Critics called the document a cover for resisting real regulation (Fortune). The same day, The New Yorker published an extensive investigative piece on Sam Altman, based on more than a hundred interviews.

Banks called to an emergency meeting over Mythos

According to Bloomberg, US Treasury Secretary Scott Bessent and Fed Chair Jerome Powell hastily summoned the heads of systemically important banks on April 7. They wanted to make sure the banks understood the risks posed by models like Mythos and were protecting their systems (Bloomberg Law).

The news spread on April 10, and according to further reports, the Canadian and British central banks are also preparing similar talks with banks (Claims Journal).

xAI sues Colorado (April 9): Musk’s company filed a lawsuit against the Colorado AI Act, which from June 30 is meant to prevent discrimination in automated decision-making. It argues that developing a model is expression protected by the First Amendment of the US Constitution (Jurist).
Florida investigates OpenAI (April 9): Attorney General James Uthmeier opened an investigation into the company over child safety, a possible link between ChatGPT and the shooting at Florida State University, and national security concerns. OpenAI pledged to cooperate (Bloomberg Government).

Security and Incidents

Supply chain attack hits OpenAI’s Mac apps

On April 10, OpenAI acknowledged that a malicious version of the popular Axios library had made its way into the signing process for its macOS apps. It was a so-called supply chain attack, in which an attacker compromises a component used by many other programs (BleepingComputer).

The malicious code had access to the signing certificate for the ChatGPT Desktop, Codex and Atlas apps. OpenAI revoked the certificate and released new versions; according to the company, user data was not at risk (Axios). Google’s security team attributed the attack to the North Korean group UNC1069.

Molotov cocktail thrown at Sam Altman’s home

Overnight into April 10, an unknown man threw a Molotov cocktail at the gate of OpenAI CEO Sam Altman‘s home in San Francisco and also made threats at the company’s headquarters. No one was injured; police arrested a 20-year-old suspect who, according to later reports, was motivated by opposition to artificial intelligence (WDBJ7 / AP).

Altman then called for a calmer tone in the heated debate about AI. On April 12, police were also investigating a shooting near the same house; whether he was the target remained unclear (Gigazine).