Week 2026-23

Jun 18, 2026

Today, a fable from Anthropic and news from Apple. @vlkodotnet

Week’s Highlight: The Short Story of Fable 5

Apple gave it their best shot last week, but the week’s highlight was stolen by Anthropic, which released the Fable 5 model — essentially the legendary Mythos 5 model, but with strict input controls. Fable 5 is 2x more expensive than Opus 4.8 and beats it across every benchmark. That means, after a long while, we finally have an AI model where the improvement is genuinely noticeable. You’ll notice this especially on tasks that take hours or even days to complete, such as vulnerability detection and deep code analysis.

So how does the model achieve strict input control? By redirecting you to Opus 4.8 whenever its control layer decides your prompt warrants it. And often a single questionable word is enough for the task to get handled by Opus 4.8 instead. With the API it’s more straightforward — there they simply reject your request. On top of that, Anthropic introduced so-called “silent” throttling. It’s invisible to the user and serves mainly as protection against model distillation, quietly altering your prompt. Anthropic promised to do this more transparently, but the suspicion lingered.

That said, the model can surprise you and get genuinely creative when completing a task — go ahead and read for yourself how it handled not having access to a keyboard and mouse.

I won’t go deeper into the evaluation, because 3 days after the release came the announcement that Fable 5 had been pulled by the US government.

The reason for the takedown was likely Amazon’s announcement that they’d found a way to bypass its built-in protection.

The fact that it happened to Anthropic of all companies is a bit of karma. Anthropic is precisely the one that has spent recent months regularly traveling to Washington to advocate for tighter controls on AI models. In the end, the US president signed an executive order (issued two weeks before Fable 5 was blocked) under which the US government will review the risk of releasing new models within a 30-day trial window.

For those of us outside the US, though, a time is coming when we may not get access to the very best AI models. There’s the option of using open Chinese models, but those may not be around forever either. China could also decide to keep them to itself, leaving API access as the only way to reach the top models.

By the way, heavily protected models are driving some interesting techniques for evading malware analysis. The latest ones insert text about biological and nuclear threats into their code, forcing a switch to a less capable version of the model.

Apple WWDC 2026

After a year, Apple finally unveiled Siri AI, built in partnership with Google. It’s based on Google’s models, which Apple then retrained. According to tests, they’re less chatty, more to the point, and — most importantly — they are not sycophantic. Alongside the server models, which come in a base and a premium version, there will also be local models. Again, there are two of them, and the better one will need 12 GB of RAM, so the latest iPhone. Apple also introduced a new Spotlight index, where information will be stored locally to enrich context, and an App Toolbox that can call into apps.

For your Macs, macOS 27 is out, integrating the new Siri AI in particular, and all Siri features will be available in the EU too. Foreign-language support is weaker, though, which is odd in itself given that Apple’s AI models are based on Google’s.

iOS 27 also has Siri AI, except most of the Siri AI functionality won’t be available in the EU. Apple blames the DMA and claims it’s all for the customer’s own good. If you’re outside the EU and China and speak fluent English, you’ll get the full Siri and everything it offers. EU customers interested in AI will probably be forced to migrate to Android.

The first tests of Siri AI are already out, and it looks quite usable.

Those are the things that caught my eye. If that’s not enough for you, I’d recommend this roundup.

By the way, I was also intrigued by the option to build Safari extensions using AI.

Security Insights

npm version 12 brings improvements that prevent potential supply chain attacks.

Microsoft is patching two 0-day vulnerabilities, but the more interesting story is the conflict with researcher Nightmare Eclipse, who claims Microsoft cheated him and has since been publishing 0-day vulnerabilities as a form of protest. These vulnerabilities go public the moment he discovers them, leaving Microsoft no time to prepare the necessary fix.

Oracle’s PeopleSoft has a publicly disclosed 0-day vulnerability with no patch available — only a configuration change to defend against it. A large amount of data has already leaked from various systems, which is rather sad, since PeopleSoft mostly manages data about people.

BIZ Insights

SpaceX’s IPO made Elon Musk the world’s first trillionaire. That’s a one with 12 zeros. SpaceX thus became a publicly traded company, which means we’ll find out every quarter how well they’re doing.

A German court ruled that, in the case of claims shown in AI Overview, Google is liable for false statements. Google argued that it’s merely displaying information from search, but the court held that it generates new content. Analyses show that AI Overview gets it wrong in roughly 9% of answers, or cites incorrect source links. At the scale Google Search is used, that’s millions of erroneous answers every day.

OpenAI, too, has now filed a confidential submission for an IPO.

Word also came out that OpenAI is negotiating with the US government over a government ownership stake in OpenAI. The government would receive the shares as a gift, and they’d be part of a broader fund holding stakes in AI companies.

AI Insights

Xiaomi unveiled the MiMo V2.5 Pro UltraSpeed model, which is ultra fast and can generate output at 1000 tokens per second. It’s a 1-trillion-parameter model with a context window of up to 1 million tokens. Its speed comes down to diffusion technology. The MiMo V2.5 Pro is available on HuggingFace.

To help you make better use of the model, Xiaomi offers MiMo Code, an agentic environment that lets you put the model to full use.

Kimi K2.7 Code is also out.

If you use GLM, then GLM 5.2 brings 1 million input tokens as well as the option to configure Reasoning modes. Available only through their paid API.

Baking diffusion functionality into models is going to be a fairly near-term trend. DiffusionGemma confirms it, speeding up local inference by up to 4x.

Finally, there’s the new OpenCV 5, the open-source library used to let computers see — that is, to handle image input processing.

.NET Insights

.NET 11 Preview 5 brings mainly news for Blazor, improvements to the new union types, and Closed class hierarchies — an interesting declaration that works within a single assembly and then lets a switch expression handle all of its derived types.

.NET 11 also improves how we launch other processes.

Links Drop

Microsoft is ending support for Office 2019 on macOS. From July 13, you’ll still be able to open a file in the program, but without the ability to edit, save, or create a new one. The recommended move is to upgrade to Office 2024 or a Microsoft 365 subscription.

Microsoft is bringing some handy improvements, too. WSL Containers is a way to no longer need Docker. On top of that, WSL Containers offer a way to get a secure sandbox environment on Windows, since you’ll be able to integrate them into your own applications.

Euro-Office is a set of products that lets you build a Google Workspace / Microsoft 365 alternative on your own servers.

The kage tool can create a copy of a website by using headless Chrome to render the page’s content and save a copy of it. You end up with a copy of the page without any JavaScript.