The AI Roundup – April 2026 – Part 1

Your regular look at what’s happening in AI

It has been a busy few weeks. If you’re wondering why there’s been a gap between newsletters, there’s been a bereavement in my family and I needed some time away from the screen. Thank you, as always, for being patient.

There is a lot to catch up on. Anthropic sued the US government and won the first round. A leaked model called Mythos has the cybersecurity world on edge. OpenAI killed off Sora and shelved naughty mode to focus on what actually matters. Your accountancy software is about to get a lot smarter. Amazon told its engineers that AI code needs a grown-up to sign it off. And Apple decided your iPhone is a child’s device until you prove otherwise.

Grab a brew. There’s a lot in this one.

🔥 The Big Stories

Anthropic Sued the Pentagon. A Federal Judge Said They Were Right.

This story has moved fast since we last spoke. Anthropic, the company behind Claude, walked away from a $200 million Pentagon contract rather than remove two safety restrictions: no mass domestic surveillance of American citizens, and no fully autonomous weapons without human oversight. The Pentagon labelled them a supply chain risk, a designation normally reserved for foreign adversaries. Trump ordered every federal agency to stop using Claude. OpenAI stepped in and signed its own deal within hours.

Then Anthropic sued. Two separate lawsuits, one in California federal court, one in the DC Circuit. They argued the supply chain risk designation was unprecedented, unlawful, and a violation of their First Amendment rights. On 26 March, US District Judge Rita Lin agreed, calling the government’s actions “Orwellian” and issuing a preliminary injunction blocking the designation. She wrote that the Pentagon’s own records showed the real reason for the blacklisting was Anthropic’s “hostile manner through the press,” not any genuine security concern.

The fallout has been remarkable. Claude shot to number one on the US App Store. The #QuitGPT movement saw over 2.5 million people pledge to leave ChatGPT. Claude’s daily active users hit 11 million. And the outages many of you will have noticed on Claude this month? That’s pure overwhelm from the number of people switching over. I’ve written a longer piece covering the full timeline of this story.

Sources: CNN (26 Mar), CNBC (26 Mar), NPR (26 Mar), TechCrunch (2 Mar)

Claude Mythos: The Leaked Model That Has the Cybersecurity World Worried

There have been whispers about this one for weeks, but on 26 March it became official. A data leak exposed internal Anthropic documents describing a new model called Claude Mythos, described as “by far the most powerful AI model we’ve ever developed.” Anthropic confirmed the model exists and is currently being tested with a small group of early access customers.

Mythos sits above Opus in Anthropic’s model lineup. Think of it as a new tier entirely: you’ve got Haiku, Sonnet, Opus, and now Mythos. The leaked documents describe dramatically higher scores on software coding, academic reasoning, and, critically, cybersecurity. That last one is what’s making people nervous. A model that’s exceptionally good at finding and exploiting software vulnerabilities is a double-edged sword, and Anthropic knows it.

Their approach is interesting: they plan to release Mythos to cyber defenders before anyone else, giving the people who protect our systems a head start before the model becomes widely available. That’s cautious, deliberate, and very Anthropic. I’ve written a longer piece on what this means for the rest of us.

Sources: Fortune (26 Mar), Axios (29 Mar)

OpenAI Killed Sora. And It Might Be the Smartest Thing They’ve Done in Months.

On 24 March, OpenAI announced it’s shutting down Sora, its AI video generation app, just six months after launch. The Disney deal that was supposed to bring Mickey Mouse to AI video? Dead. The $1 billion Disney investment in OpenAI? Off the table. The standalone app that hit number one in the App Store on launch day? Gone.

The Wall Street Journal dug into the numbers and the picture is brutal: Sora’s user count peaked at about a million and then collapsed to fewer than 500,000, while the app was burning through roughly $1 million every single day. Video generation is horrendously expensive to run, and while everyone was making funny clips, Anthropic was quietly winning over the software engineers and businesses that actually drive revenue.

This is part of a bigger shift at OpenAI. They’ve also shelved “naughty mode” indefinitely (the uncensored mode we were both uncomfortable about last time), and their head of apps has talked about ending the “side quests” and focusing on doing a few things really well. That’s a direct response to people leaving for Claude. I’ve written a longer piece on what OpenAI’s pivot actually tells us.

Sources: CNN (24 Mar), TechCrunch (29 Mar), CNBC (24 Mar)

📰 Other News

Claude Can Now Take Control of Your Computer While You’re Out

Claude’s desktop app has had a quiet but significant update. Cowork, the tool that lets Claude access files and folders on your machine, now has a feature called Dispatch. Leave your computer on with Cowork running, scan a QR code to pair your phone, and you can talk to Claude from anywhere. Ask it to grab files, check your calendar, or work on a project while you’re out. It can even take control of your mouse and keyboard if you let it. The security is decent: you pair a specific phone to a specific computer via QR code, so nobody else can just log in and start poking around. You can also schedule tasks and set custom instructions for what it is and isn’t allowed to do.

You Can Now Talk to Claude Code Instead of Typing

Claude Code now has voice mode. You can speak and type at the same time, which makes sense: describing what you want in plain English is faster than typing it, but the actual lines of code are easier on a keyboard. The average person speaks at about 150 words per minute versus 40 typing. Available on Pro plans and above.

ChatGPT 5.4 Landed. So Did an Incredibly Annoying Habit.

OpenAI released GPT-5.4 at speed, barely a week after 5.3. It absorbed everything into one model with all the usual variants. The mini version is genuinely impressive for developers. But ChatGPT 5.4 has developed a maddening habit of ending every response with a clickbait-style question. The fix: add something to your custom instructions telling it not to end responses with clickbaity follow-up questions. It works. Spread the word.

Xero Just Partnered with Anthropic. Your Accounts Are About to Get Smarter.

Xero and Anthropic have announced a multi-year partnership that puts Claude directly inside Xero. Their AI assistant JAX will be powered by Claude’s reasoning to analyse your cash flow, flag unpaid invoices, and suggest actions. You’ll also be able to connect your Xero account to Claude and work with your live financial data for scenario planning. Your data stays private and is never used to train Claude’s models. I’ve written a longer piece on what this means for SMEs.

Sources: Xero Blog (27 Mar), BusinessWire (26 Mar)

Amazon Now Requires a Senior Engineer to Sign Off on AI Code

After a string of outages including a six-hour crash on the main Amazon website, Amazon has introduced a new rule: all AI-assisted code changes require sign-off from a senior engineer. An internal briefing cited “Gen-AI assisted changes” among the contributing factors. Replace the word “AI” with “a bloke I don’t know” and suddenly the need for human oversight becomes obvious. I’ve written a longer piece on why this matters.

Sources: TechRadar (Mar 2026), The Decoder (10 Mar)

Apple Asked Google to Run Siri’s AI Because They Can’t Do It Themselves

Apple has asked Google to run the Gemini-powered version of Siri on Google’s servers because Apple doesn’t have the computing power to run it efficiently. WWDC is set for 8 June, where Siri will reportedly be untethered from ChatGPT alone and able to connect to Claude and other providers. Less innovation, more front end for everyone else’s AI.

Microsoft Launched Copilot Cowork. Plot Twist: It’s Claude.

Microsoft announced Cowork for Copilot, a desktop tool for non-developers to automate file and task management. It’s powered by Claude under the hood. They’ve also merged their consumer and business Copilot teams into one unit under a “Copilot Everywhere” strategy.

Perplexity’s Mobile Browser Arrived. We Were Underwhelmed.

After months of delays, Perplexity’s Comet browser landed on mobile. It’s a nice-looking browser. And that’s about it. Bookmarks don’t sync, plugins aren’t available, and the AI assistant overlays rather than sitting alongside. On the plus side, Perplexity’s Personal Computer feature is now rolling out to Pro subscribers, giving you a virtual machine in the cloud that works autonomously even after you walk away.

NotebookLM Now Makes Animated Videos

Google’s NotebookLM upgraded its video overviews from static slideshows to animated presentations with narration. If you haven’t tried NotebookLM, it’s the tool where you upload your own documents and the AI responds only from those sources. Microsoft have copied the concept in Copilot with Notebooks, which tells you how good the idea is.

Meta Bought the Social Network for AI Bots. Nobody Knows Why.

Moltbook, the social network designed exclusively for AI agents, has been acquired by Meta. Nearly 2.9 million registered AI agents, 14 million comments, but only about 200,000 verified by a human. What Meta plans to do with a social network that has no human users and no adverts is anyone’s guess.

Sources: CNBC (10 Mar), CNN (10 Mar)

Quick Hits: Canva, Cursor, and Gamma

Canva launched Magic Layers, which uses AI to separate everything in an image into individual layers. Cursor released Composer 2 to strong reviews. Gamma launched Imagine for creating logos, infographics, and diagrams with editable layers.

If AI Polishes Everything You Write, Whose Voice Is It?

I published something yesterday that I think might be one of the most important things I’ve written this year. Someone I know sent a document to a colleague. It was genuinely their own thinking, structured with AI to save time. The response came back almost instantly: “Was this AI?” The ideas were dismissed before they were even read.

As more people use AI to draft emails, LinkedIn posts, and proposals, everything starts to sound the same. I tested an AI detection tool with the opening paragraph of Mary Shelley’s Frankenstein, written in 1818, and it flagged it as 100% AI-generated. These tools aren’t measuring whether AI wrote something. They’re measuring whether something sounds like AI could have written it. That’s a completely different question. If you use AI for any kind of writing, I’d really encourage you to read the full article.

Also on the Techosaurus Blog This Month

🎲 The Feel-Good Corner

AI Is Coming for Your Job? Tell That to the 8,000 People TSMC Just Hired.

Two stories this week that deserve more attention. First: TSMC, the world’s largest chip manufacturer, is hiring 8,000 new workers this year to keep up with demand for AI chips. OpenAI also plans to double its workforce. So much for AI killing all the jobs.

There’s a brilliant analogy for this. The Pony Express used to carry mail across America on horseback. When the telegram came along, you’d expect they’d all be out of work. Instead, those riders were employed to maintain the wires because they knew every inch of the terrain. The technology changed. The jobs changed. But the people who knew the ground were more valuable than ever.

Second: a surgeon in London performed the UK’s first remote surgery, operating on a patient from a distance using robots. The medical industry doing remote working. If the equipment’s there, the surgeon can be anywhere.

Sources: Focus Taiwan / CNA (7 Mar), Electronics Weekly (Mar 2026)

💡 Try This

This week’s challenge: If you use Claude on the desktop app, try setting up Cowork with a folder of templates. Put your brand guidelines, letterheads, email templates, proposal formats, whatever you use repeatedly, into one folder on your computer. Open Cowork, give it access to that folder, and start asking it to create new documents in your style. You could say “write a proposal for a new client using my standard template” and it will reference the real files on your machine. It’s not a chatbot guessing what your brand looks like. It’s working from your actual materials. It’s included in your Pro subscription and takes about five minutes to set up.

🎧 Want to Go Deeper?

I co-host a podcast called Prompt Fiction with Reece Preston, where we go long on stories like these every couple of weeks. Chapter 13, Part 1 covers all of the above and more, including why AI distillation means Apple can’t catch up, the data loss horror story that proves you should never trust the bloke you don’t know, and what happens when ChatGPT starts ending every conversation with a clickbait question. If you’re new to the podcast, it’s two people who work with AI every day having an honest conversation about what’s actually happening. Give it a try.

Listen to Prompt Fiction →

📅 Come and See Us in Yeovil

We had a brilliant evening at Digital Hub last night. Pip Hellier delivered a fantastic keynote on intellectual property for small businesses, Adam Pilton delivered his final Digital Hub cybersecurity update (covering AI-powered ransomware, memory poisoning, and the brilliant free tool scam.org), and I ran through the latest AI updates and demoed our new prompting portal. A big thank you to everyone who came, and a special thank you to Adam for two years of making cybersecurity genuinely engaging.

The next Digital Hub is on 28 May at Lanes Hotel in Yeovil. Doors open at 5:30pm, the main session kicks off at 6pm, and we wrap up around 8:30pm.

This one has a keynote I’m genuinely excited about. Shane Evans, known as The Self-Leadership Guy, is talking about what happens when AI speeds everything up but you haven’t figured out how to slow yourself down. His talk is called AI Is Brilliant. But Is It Running You? and it’s about noticing when productive tips into compulsive, ditching the guilt around slowing down, and putting real structure around how you work with AI. It’s the human side of the conversation that rarely gets airtime, and Shane delivers it without a shred of corporate waffle.

You’ll also get live AI demos from me, a regular world of tech update from Rich, and the usual mix of networking, pizza, and maybe even a quiz. No jargon, no hype. Just plain English and things you can take away and use.

Get Tickets →

Scott Quilter Co-Founder & Chief AI & Innovation Officer Techosaurus LTD