Here are three things I found interesting in the world of AI in the last week.

Anthropic's best model is also its most restricted

Anthropic shipped Claude Fable 5 on Tuesday. It's essentially Mythos (the model which found all the security vulnerabilities) with some extra safeguards. If you ask it anything about cyber security or biology it will restrict access and automatically downgrade to Opus.

I've been using it non stop since launch and it's the best model available by a long way. Anthropic have included it in their pro, team and enterprise plans until June 22 and then you'll only be able to access it through extra usage (i.e. pay as you go api tokens). The price is twice that of Opus so, not cheap.

Anthropic have said publicly they will be putting it back in the plans when they get more compute. But there is no hint about when that might be, so maybe don't get too attached unless you're willing to open your wallet.

Two other things of note:

First, there's no zero data retention. Every prompt and output is held for 30 days, and anything flagged as policy-violating can sit around for up to two years. That overrides existing enterprise ZDR agreements, and it's the only Claude model that can't run under them. Anthropic say it's for safety, not training, but Microsoft has already restricted internal access while their lawyers look at it. I've seen claims that using Fable could be viewed as violating NDAs.

Second, and this is the controversial one, a paragraph buried deep in the 319-page system card admits Fable 5 silently degrades its answers when you ask it about cutting-edge AI development, like building the infrastructure used to train large models. The cyber and bio safety fallbacks at least tell you when they trigger. This one doesn't. It just gets quietly worse and you'd never know.

After a wave of backlash they grudgingly made the throttling visible so AI researchers now get told when they're being downgraded instead of being silently handed a worse model. Took a public roasting to get there, but credit where it's due.

So the only AI researchers who can use Fable to speed up their work are the ones at Anthropic, everyone else gets a dumber version when asking about AI research. Anthropic claim it is for safety and de-accelerating AI development, but it looks a lot like de-accelerating everyone else, not themselves.

I guess if you say safety enough you can justify anything you want.

If you view Anthropic as the most trustworthy gatekeepers for stewarding super intelligence into the world responsibly then that could be a good thing. If you are concerned about a soon to IPO, US megacorp using super intelligence to dominate an emerging market then maybe you don't like that idea.

Either way, Fable/Mythos is more than an incremental improvement and it unlocks a whole lot of potential. For the right people. And the right price.

In a final twist, the US government forced Anthropic to block access to all foreign nationals, so bye bye fable access for now.

Apple gave up building its own AI and rented Google's

At WWDC on Monday Apple announced "Siri AI", a full rebuild of the assistant. Finally, a non sucky Siri! Again!

The new bit, is that it runs on a custom version of Google's Gemini, reportedly for around a billion dollars a year. The standalone Siri app, the cross-app context, the multi-step requests, all the stuff they promised in 2024 and never shipped, is now powered by a competitor's model.

This is a big concession from a company that markets itself on owning the whole stack. Apple tried to build its own foundation models, the personal Siri slipped for over a year, and in the end they decided renting was better than shipping late again. I actually think it's the right call. Apple's edge was never going to be the model, it's the billion devices and the privacy story around them. Better to put a good model in front of users now than wait another year for a worse in-house one.

The reception was rough though. The stock dropped on the day, analysts called the updates "lacking", and Gene Munster reckons real availability slips to mid-2027. It's also blocked in the EU over the Digital Markets Act and unavailable in China. So the demo looked fine, but the question of whether Apple can actually ship AI on time is still wide open.

But still I have this niggling feeling that not getting into the "spend billions and billions on AI data centres and foundational models game" might look like a smart pay when the AI bubble starts to lose steam.

Even Google can't pay for this out of pocket

Alphabet raised about $85 billion in early June, upsized from an initial $80bn. This is equity, not debt, including a $10bn private placement from Berkshire Hathaway. The money is for AI infrastructure: GPU clusters, data centres, and power deals. Alphabet's 2026 capex guidance is now $180 to $190bn.

The thing is, Google has more cash than almost anyone. They make a ridiculous amount of profit - and still chose to dilute shareholders to fund the buildout. This is a company that was historically valued for being capital-efficient, and it's now stacking an $85bn equity raise on top of more than $100bn in debt. When one of the most profitable businesses in tech decides it can't pay for AI out of cash flow, that tells you something about the scale of the bet.

With the SpaceX, Anthropic and OpenAI IPOs on the horizon the frontier labs are now looking at raising something like 300 - 400 billion of new money which is pretty unprecedented. I reckon Google are getting in early as this would be as likely a time as any for investors to get cold feet about the viability of the whole mega-spend. If any of the big IPOs fail and we see a bubble pop then they are selling at the top. If things keep rolling then they've got the cash to keep pace.

The surprise raise also means there is a little(?) less free cash out there for the other labs to bid for, which I'm guessing was also part of the calculus.

So, it's popcorn time for us regular folks while the big kids are putting their pocket money on the poker table.