This week in AI hit different.

Both Anthropic and OpenAI are dropping major desktop overhauls that could reshape how teams build and run agents. Meta came back from the dead with a model nobody saw coming. Microsoft declared an internal emergency over Copilot. Sam Altman's home got attacked. Twice.

Oh. And Anthropic built a model so dangerous they won't let anyone near it.

The kind of week where missing the news means falling behind fast.

Here's what matters.

Reports from TestingCatalog reveal Anthropic is preparing a major overhaul of Claude Code Desktop, codenamed Epitaxy, potentially shipping this week.

The redesign brings a single-window layout with panels for plans, sub-agent tasks, and live code previews. Multi-repo support included.

Anthropic is also building Coordinator Mode, which would let Claude orchestrate work across multiple parallel sub-agents while staying focused on high-level planning. Users will reportedly create custom agents directly inside the app on the fly.

On the OpenAI side, multiple employees confirmed major Codex updates are coming this week too. The company is testing a Scratchpad feature for queuing parallel tasks from a to-do view, a clear signal that Codex is expanding beyond coding into a broader AI work hub.

OpenAI also appears to be consolidating ChatGPT, the Atlas browser, and its agentic tools into one super app with Codex as the central platform. Snowflake emojis from OpenAI employees sparked speculation about a model codenamed Glacier, believed by some to be GPT-5.5, potentially pairing with the launch.

What it means: Two major AI companies racing to own your desktop. Whichever one ships a seamless, unified agent experience first this week pulls ahead with power users and dev teams. What actually drops matters. Watch closely.

What if shipping a production-ready AI agent didn't require months of infrastructure work?

Anthropic launched Claude Managed Agents in public beta this past week. Full production stack. Build and deploy cloud-hosted agents at scale without managing the underlying infrastructure yourself.

Sandboxed code execution, credential management, scoped permissions, checkpointing, end-to-end tracing. Anthropic handles all of it. Teams just define the task, tools, and guardrails.

Long-running sessions persist through disconnections. Multi-agent coordination lets one agent spin up and manage others. MCP connectivity means plugging into your existing software stack is relatively painless.

One important heads-up. This lives on Anthropic's developer platform, not claude.ai. Billing is usage-based via API, and parallel sub-agents add up fast. Budget ahead before you build.

What it means: The biggest barrier to building production agents just got meaningfully lower. If your team has an automation use case but lacks the infrastructure muscle to execute it, this is the most direct path yet. Just watch those token costs. They climb faster than expected.

Welp. The company building the robots just proposed taxing them.

OpenAI released a 13-page policy document titled "Industrial Policy for the Intelligence Age: Ideas to Keep People First," widely being called the Superintelligence New Deal.

The argument: AI is accelerating so fast that The US may need a new social contract, comparable to the Progressive Era and the New Deal, to get ahead of mass job displacement, cyberattacks, and social instability.

Proposals include a national public wealth fund seeded partly by AI companies, taxes on automated labor, a four-day 32-hour work week that returns AI productivity gains directly to workers, and treating AI access as a basic right for workers and schools.

The document also calls for containment plans for dangerous autonomous AI and automatic safety net expansions when AI-driven disruption hits preset thresholds.

Most of this will never become federal law. States are the more realistic watch, especially California. And yes, it is a little awkward to call your own product a basic human right while openly acknowledging it will displace millions of workers.

What it means: The robot tax and public wealth fund concepts will resurface at the state level. Think of this as the framework for policy debates that are coming regardless. If you're mapping long-term AI strategy, start modeling now what a labor displacement tax environment does to your automation ROI.

Nobody had Meta in the top three AI models heading into this week.

Yuuuuup.

Meta announced Muse Spark, the first model from the newly formed Meta Superintelligence Labs. Built from the ground up over more than a year, backed by a $14.3 billion deal to bring Alexandr Wang from Scale AI as Chief AI Officer. Developed internally under the codename Avocado.

Now live on Meta's AI app and website. API access coming soon. Proprietary, not open source. Currently free.

According to Artificial Analysis, Muse Spark already matches top models from Google, OpenAI, and Anthropic on language and visual tasks, though it falls behind on coding and abstract reasoning. It currently sits third on Arena, the blind AI preference leaderboard. Contemplating mode, which orchestrates multiple parallel agents for complex reasoning, is rolling out gradually.

Real-world performance on highly complex tasks is still mixed. But landing in the global top three with the first model in a completely rebuilt series is well ahead of where most people expected Meta to land.

What it means: Meta is legitimately back in the AI conversation. Muse Spark isn't the top model, but this level of early competitiveness from a rebuilt-from-scratch lab is a real signal. If you've written Meta off as an AI also-ran, it's time to revisit that take.

According to BNP Paribas analyst Stefan Slowinski, Microsoft CEO Satya Nadella has declared a Copilot Code Red inside the company, an all-out push to fix Copilot's performance and user experience.

Investor frustration has been building. Copilot's traction remains limited despite Microsoft's dominant enterprise position, and competition from Anthropic is cited as a primary driver of the urgency.

Nadella's initiative reportedly includes the E7 suite launching around May 1, with ongoing updates throughout the year. Early product feedback is reportedly improving. Slowinski also noted Azure could still outperform expectations due to growing token demand and higher GPU pricing.

Nah, Copilot disappointing enterprise users isn't new news. Many companies that invested heavily in 2023 and 2024 found it underwhelming as Anthropic and OpenAI pulled ahead on quality. The Code Red is an acknowledgment of exactly that. But Microsoft is the largest investor in both OpenAI and Anthropic, so even users who defect to those platforms still generate returns for Microsoft. And Copilot's embedded access across enterprise systems is an advantage nobody else can replicate yet.

What it means: The E7 suite launching May 1 is the real test. If it delivers, Copilot's distribution advantage at enterprise scale becomes formidable. If it underwhelms, expect accelerated flight to Claude and ChatGPT. Watch that launch closely.

San Francisco police confirmed two separate attacks on OpenAI CEO Sam Altman's home in the span of four days.

First: a 20-year-old man from Texas allegedly threw a Molotov cocktail at Altman's property early Friday morning. Security extinguished the fire. No injuries in either incident. The suspect was arrested an hour later outside OpenAI's headquarters, allegedly threatening to burn it down.

Then early Sunday, suspects in a car allegedly fired shots at Altman's property before fleeing. Police traced the vehicle via surveillance and arrested two suspects. Three firearms found at their residence. Both booked for negligent discharge of a firearm.

Altman published a blog post after Friday's attack, sharing a photo of his family and calling for de-escalation of rhetoric in the AI industry.

Silicon Valley is a bubble. The rest of The US is watching AI eliminate jobs in real time, and unlike the internet, which reshaped the economy over decades, AI is doing it in years. The executives running these labs have very visible faces. That anger has to land somewhere.

What it means: This is not a one-off. As AI-driven displacement accelerates, the people most publicly associated with it face growing personal risk. Physical security needs to be a core concern for AI leaders right now. This week was a preview of a longer and uglier trend.

Sheesh.

Anthropic announced Claude Mythos Preview this past week, a new frontier model so capable at finding and exploiting software vulnerabilities that the company is keeping it private.

In testing, Mythos found thousands of critical vulnerabilities across major operating systems and web browsers, including a 27-year-old flaw in OpenBSD and a 16-year-old bug in FFmpeg that years of human security research never caught.

Rather than a public release, Anthropic launched Project Glasswing, restricting Mythos access to partners including Apple, AWS, Google, NVIDIA, Microsoft, Cisco, and more than 40 other organizations, to use it defensively before a model this capable reaches the open market.

This is the first time in the modern AI era a major frontier model has been withheld from the public and released only to elite partners.

Anthropic says Mythos wasn't trained as a cybersecurity weapon. Its advanced coding capabilities just happened to surface vulnerability-finding abilities even top human experts had missed. Worth noting too: Anthropic's IPO is expected later this year, and a major source code leak landed just weeks prior.

What it means: When the same model can both defend critical infrastructure and potentially destroy it, the public access calculus changes completely. Pay close attention to who gets access to next-generation frontier models and who doesn't. That gap is about to widen fast.