Hello fellow keepers of numbers,

It was a glorious 48 hours with Fable 5. It almost finished all the work I’ll ever have in my lifetime before it was taken away. So close.

But in all seriousness, the oversight and regulation is coming for AI sooner rather than later. Anthropic was told to limit access to Fable 5 due to potential national security risks. In other news, SpaceX officially exercised their option to acquire Cursor. Might not seem that relevant to accounting on the surface, but it could be a competitor to Claude Code, Claude Cowork, and Codex.

Plus, a few weeks ago, I included a demo on how to use Codex on your phone. Now, I’ve included a demo on how to use Dispatch in Claude, which lets you take actions from your phone as well.

THE LATEST

The US government ordered Anthropic to cut off Claude Fable 5 and Mythos 5 for all foreign nationals, citing national security. Anthropic, which cannot verify a user's nationality in real time, complied by disabling both models worldwide for every customer, just after launching Fable 5.

Anthropic says the order stems from a "narrow potential jailbreak," a method of getting the model to read a codebase and fix software flaws, and it disputes the move, saying the discovery should not "be cause for recalling a commercial model deployed to hundreds of millions of people." Semafor reported that the order followed White House concern over a suspected China-linked group accessing Mythos, which Anthropic disputes.

All other Anthropic models, including Opus 4.8, stayed online. Anthropic says it is working to restore Fable 5 and Mythos 5 "as soon as possible" but has not given a timeline.

Why it’s important for us:

And as quickly as it came, it vanished. It was a special 48 hours.

If anyone was skeptical about AI being a national security risk, that skepticism is gone. We’ve been headed for more oversight and regulation for a while now, but I didn’t think it’d come in this fashion. Even the AI companies have been asking for regulation lately, including in a blog post from Anthropic where they suggested a worldwide pause. This probably wasn’t what they had in mind.

So what actually changes for us? Probably not much in how we use Claude models day-to-day. It’s unlikely Fable 5 was required for most accounting-related tasks. But there are a few important factors we should consider moving forward.

(1) The order by the US government specifically limited access to only US citizens. This excludes even foreign nationals working inside the US, many of whom work for the AI labs. We could be headed toward a world where US citizens have access to different models than non-citizens. This could impact offshore teams and remote workers in accounting firms.

(2) It may impact how we roll out new models into existing agents and automations. If a new model can be pulled at any moment, it might be best to continue using older models in those agents and automations until we’re confident the new model is going to stay. If you wire up an agent with a new model that gets pulled, your agent could break if it doesn’t automatically default to another model.

(3) It’s likely going to impact vendors. They’re using these models behind the scenes for their AI features. Hopefully, something like this wouldn’t really impact end users. But pulling the rug in this fashion could disrupt operations, even if just for a few hours.

SpaceX said it has exercised an option to acquire Cursor, the AI coding tool, in an all-stock deal valued at $60 billion. The option let it either pay roughly $10 billion for a partnership with Cursor or buy the company outright for $60 billion.

The company said its goal is "building the world's most useful AI models," and that it has spent the past few months jointly training a model with Cursor that will be released in Cursor and Grok Build, xAI's coding agent. SpaceX owns xAI, the maker of Grok, and now folds Cursor into that stack.

Reportedly, the deal will help SpaceX compete with OpenAI and Anthropic, which have their own AI coding tools. The acquisition is expected to close later this year.

Why it’s important for us:

What does a rocket company have to do with accounting? Nothing. What does an AI coding tool have to do with accounting? Mostly nothing. But put them together? Nope, still nothing. Let me speculate for a minute.

Cursor has actually been a category leader for agentic AI for over a year. They were one of the first tools that moved to an agent-first interface, similar to Claude Cowork and Codex. But they were heavily focused on developers writing code. According to the SpaceX tweet announcing the deal, the goal is “building the world’s most useful AI models.” Emphasis on “useful.” They could’ve said “powerful” or “intelligent.” But they chose “useful.”

I think this is a hint at the market they’re going after next. Just like OpenAI, Elon has seen the waves Claude Cowork has made in the enterprise space for non-developers. Cursor is in a really strong position to tackle this market.

So, why do we need another Claude Cowork or Codex? [when is this guy going to stop asking himself questions?]

Right now, Cursor is model-agnostic. That could potentially change if SpaceXAI launches new models. But I suspect, even if they do, Cursor will remain model-agnostic. You get ChatGPT, Claude, Gemini, several other open-source models, and Cursor’s own model (which is very capable).

There’s a lot of benefit in being able to use different models for different tasks where they excel. But think back to the story above. We just watched a great model get pulled overnight. In a world where the best model changes every few weeks, a tool that lets you swap between and choose different models for different tasks seems genuinely useful.

None of this is immediate, and some of this is my speculation. But this is something interesting to watch over the next few months.

TRENDING NEWS

Anthropic added artifacts to Claude Code for interactive pages like project dashboards: This has been available in Claude Cowork, and now comes to Claude Code. Interactive artifacts that can be saved and viewed at any time, which can also include live data from connectors.

OpenAI rolled out scheduled tasks in ChatGPT and will retire Pulse within 14 days: Another Claude Cowork feature makes its way into a separate tool. Scheduled tasks are automations to monitor websites or specific data.

Ignition released an MCP for connecting proposal, billing, and payment data to AI assistants: One of the most popular engagement letter softwares for accountants gets an MCP. Obligatory round of applause for the MCP release. [clapping]

Expensify launched an MCP for querying expense data through AI assistants: Obligatory round of applause for the MCP release. [clapping]

Anthropic introduced centrally managed authorization for MCP connectors in Claude, starting with Okta: This is a quality-of-life improvement for firms rolling out Claude to the team at large. Connectors can now be managed and automatically added for staff. Previously, each staff had to add the connector individually. I expect this to roll out more broadly beyond just Okta.

Pilot launched Meridian, an AI platform it says can run the full month-end close: Another month-end close launch. We had Ramp, then Digits, and now Pilot. Everyone clearly sees close work as one of the first places AI can actually take work off a firm’s plate.

Gusto released six AI agents for accounting firm business development: I’m a bit confused by this one, but I guess everyone wants in on the AI agents mania. Business development is a known weakness in the profession, but Gusto doesn’t seem like the first place I’d go for these kinds of agents.

OpenAI launched Record & Replay for Codex, turning a demonstrated recurring task into a reusable skill: I’m looking forward to testing this one. There are some vendors going after this space right now as well. If this works, it’s a very clean way to turn random admin tasks into repeatable workflows.

Anthropic updated Claude Design with design systems, canvas editing, Claude Code sync, and PowerPoint export: Claude Design is still so new. In my early testing, it seemed incredibly powerful, but it was siloed. Connecting it with the rest of the Claude stack makes a lot of sense and should make it much easier to use. Claude Design is great for website design, custom app design, and branding. It could be useful for PowerPoint, but it’s still not great at that yet.

Perplexity launched Brain, a memory system for Computer that builds a context graph across tasks: It’s been a minute since we heard from Perplexity. They’re angling for the Obsidian users who’ve built a “second brain.” This sounds like it’s intended to be automated memory with no setup required, whereas the “second brain” requires some development and organization before it’s useful.

PUT IT TO WORK

I don’t know about you, but my brain decides to be its most productive as soon as my head hits the pillow at 11pm. Why not just let Claude do my work while I lay in bed?

First, to use Dispatch, you need to enable it in your org settings. Navigate to Cowork and then toggle on “Enable Dispatch” (see below).

Setup is as simple as with Codex. On the desktop app, click the Dispatch button in the left sidebar. You should see the below.

Click “Get Started.” On the next screen, you’ll see the below.

Give Claude access to your files. Your computer must be awake in order to use Dispatch, so you can also choose to have the desktop app keep your computer awake.

On your phone, open the Claude app and click Dispatch. You should be able to use it right away, but it might prompt you to scan a QR code to finish setup.

From there, prompt it just like you would Claude Cowork.

You don’t connect specific folders like you do for Cowork tasks. It has access to your whole computer, so treat it accordingly. You need to specify exactly where your files are so it knows where to navigate (which will also save usage). This means you also need to be aware that it has access to all files on your computer, which brings in additional risks.

WEEKLY RANDOM

Midjourney makes AI images. But I guess they also make medical scanners now too? Maybe one of the great 180s in recent memory. That’s quite a leap for Midjourney.

The pitch is a head-to-toe ultrasound in 60 seconds for a few bucks. You get lowered into a shallow pool of water lined with half a million sensors, and a computer turns the soundwaves into a 3D picture of your muscle, fat, bone, and organs.

No radiation like a CT, no giant magnets like an MRI. They’ve even talked about opening a "Midjourney Spa,” so you can get your organs mapped between the sauna and the cold plunge. Maybe do it before the cold plunge though…

This is currently just a prototype, and it doesn’t yet work like they suggest the final product will. The AI imaging tools also still need to be built, ironically. It’s also not going to replace all medical scans. There are plenty of others that are much better at detecting important things than an advanced ultrasound.

Still, this is another really cool idea in the healthcare space. An MRI can take 30 minutes to a few hours and is typically really expensive. So getting close to that in just one minute for a few dollars completely changes who can afford to monitor their own health on this scale.

It seems like it becomes incredibly useful if you get scanned regularly. If it’s 60 seconds and only a few dollars, you could do it every month and possibly catch diseases and cancers much earlier than you might otherwise, even without advanced AI imaging.

It’s very early, but it’s exciting to see announcements like this that have the potential to positively impact tons of lives.