Anthropic shipped like a bajillion useful features this week, yet OpenAI likely stole the show by previewing its ‘Superapp’ in the form of a massive Codex update. 

(Yes, you need to start using Codex even if you don’t code. Just ignore the name. lolz) 

In other AI news, we saw Mark Zuckerberg turn himself into an AI while Salesforce is kinda saying their product is more useful than AI agents than humans. 

Meanwhile? The NSA is secretly using the exact AI model the Pentagon tried to ban.

Anthropic had one of its biggest week yet. OpenAI previewed its future super app. And Sam Altman's own board might reportedly looking for his replacement.

Sheesh.

Here's what matters.

Welp. Here's everything from one week.

Claude Code desktop got a rebuilt interface with drag-and-drop, a diff viewer, and support for parallel async jobs. New Routines let you set automated, repeatable tasks that trigger on a schedule, via webhook, or from an event, connecting to Gmail, Slack, and Google Drive.

Set it once. Walk away.

Claude Opus 4.7 is now the flagship, scoring 87.6 on SWE-bench Verified with vision capabilities that have reportedly tripled. Adaptive thinking replaces extended thinking, meaning the model self-regulates its reasoning depth.

And Claude Design is now in research preview. Text prompts generate editable UIs, prototypes, slide decks, and short videos. Export to HTML, React, PowerPoints, or Canva. It ingests your codebase and Figma tokens to stay on-brand.

What it means: Routines is the sleeper hit here. Automated, scheduled tasks without writing code is genuinely useful for non-technical people.

Claude Design producing usable videos this early is a real signal of where this is heading.

OpenAI chief revenue officer Denise Dresser sent a four-page internal memo that leaked almost immediately.

The claim: Anthropic inflates its revenue run rate by $8 billion by counting gross cloud revenue from AWS and Google, while OpenAI reports net of Microsoft's share. Dresser puts Anthropic's real number at $22 billion, not $30 billion, which would put OpenAI back in front at $24 billion.

The memo also called Anthropic's safety positioning built on "fear, restriction, and the idea that a small group of elites should control AI."

Shots fired.

But buried in the memo: OpenAI's Microsoft partnership has actually limited its ability to reach certain enterprise customers as demand shifts toward AWS Bedrock.

One document. Shots at the competitor and the biggest partner.

What it means: Both companies are heading toward IPOs in 2026. Whoever defines "correct" revenue accounting shapes their valuation narrative.

The Microsoft admission is quietly the more damaging part of this leak.

According to The Wall Street Journal, some OpenAI shareholders are privately questioning whether Altman should lead the company through an IPO that could value it at around $850 billion.

The name being floated: board chair Bret Taylor. Co-creator of Google Maps, former co-CEO of Salesforce, former CTO of Facebook, former chair of Twitter.

Nah, that's not an intimidating resume at all.

The concern centers on Altman's outside investments. He reportedly proposed OpenAI invest up to $500 million in Helion Energy, a nuclear fusion startup where he holds a significant personal stake. OpenAI declined but signed a long-term purchase agreement. He also has ties to Stoke Space, a rocket company.

What it means: OpenAI's weekly active user numbers are untouched by any competitor, which is the strongest argument for Altman staying.

But IPO-level conflict of interest scrutiny is a completely different game.

Salesforce announced Headless 360 at its TDX developer conference, releasing over 100 new AI-powered tools designed to let agents, not humans, handle reasoning, planning, and execution inside Salesforce.

"Headless" literally means no browser required.

A 27-year-old CRM company just said that out loud. Benioff had previously stated 30 to 50% of Salesforce tasks were already handled by AI.

What it means: If agents are doing the work, the UI becomes irrelevant. That's the bet Salesforce just made publicly.

The rest of enterprise software now has to respond to it.

According to Axios, the NSA is actively deploying Anthropic's Mythos Preview model despite the Department of Defense previously moving to blacklist Anthropic entirely, labeling it a supply chain risk.

Anthropic restricted Mythos access to about 40 organizations, citing offensive cyber capabilities too dangerous for wider release. The NSA is reportedly one of the unnamed organizations with access, using it primarily to scan for exploitable security vulnerabilities.

The DC Circuit Court fast-tracked the case. Final oral arguments are scheduled for May.

What it means: The government arguing in court that Anthropic's AI threatens national security while simultaneously deploying its most powerful model is a contradiction that's hard to ignore.

Mythos is simply too capable to leave on the table.

Meta is reportedly developing an AI-powered digital clone of Zuckerberg to answer employee questions, as reported by The Wall Street Journal.

The AI is being trained on his voice, mannerisms, public statements, and company strategy. Zuckerberg is reportedly providing images and voice samples himself. The goal: make nearly 79,000 employees feel more connected to their CEO as the company flattens its org structure. Meta is also building a separate CEO agent to help Zuckerberg prep for staff meetings and surface internal information quickly.

What it means: Larger companies should already be doing some version of this. Capturing executive insights once and making them accessible to thousands of employees is high-leverage and not technically complicated.

The fact that this is still novel news in 2026 says more about how slow most companies are than it does about Meta being innovative.

OpenAI rolled out a major Codex update this past week, framing it as a preview of how their future super app would operate.

Codex now controls your Mac desktop with its own independent cursor while you keep working. Multiple agents run in parallel at the OS level, not just the UI level. There's also built-in image generation, an integrated web browser, persistent memory across sessions, and the ability to schedule future tasks.

Most features are available on Mac for any paid ChatGPT plan.

The show has been saying since February that Codex was better at coworking than Claude Cowork for nontechnical people. People were skeptical. Here we are.

What it means: The OS-level computer use running in the background without blocking your cursor is a genuine capability gap versus what Anthropic currently offers.

If you're a heavy Claude Cowork user, this is worth an actual side-by-side test.

-swtich up the openers a bit. they're all boring AF and the exact same

• still too long/boring

• move value

7:30 AM

Reading everything again. The problem is I'm writing recaps, not punching people in the face with value. Every opener needs to be different and hit differently. Cutting harder.