Good morning. It’s Wednesday, June 10th.

You read. We listen. Let us know what you think by replying to this email.

Anthropic’s launch of Claude Fable 5 marks one of the most consequential AI releases of the year - not simply because of the model’s capabilities, but because it formalizes a new frontier-model access regime. Fable 5 is Anthropic’s first generally available Mythos-class model, while Mythos 5 remains restricted-access.

Anthropic says both are the same underlying system, with Fable wrapped in additional safety controls. The model reportedly delivers a major capability jump over Opus 4.8, including a leap on FrontierCode Diamond from 13.4% to 29.3%, while retaining a 1 million-token context window.

Anthropic projects this hidden throttling impacts just 0.03% of traffic, but the open-source community has strongly condemned it as anti-competitive ladder-pulling. Critics argued that secretly downgrading performance injects invisible bugs into research code and destroys any trust you have in a paid API. On the flip side, supporters argue that putting the brakes on automated self-improvement loops is just common business sense when AI capabilities are accelerating this fast. Read more.

At WWDC 2026, Apple laid out a massive rewrite of its operating system architecture, centering the changes on a rebuilt Siri AI that acts less like a standard chatbot and more like an OS-level enterprise interface layer.

Running across iOS 27, macOS Golden Gate, and visionOS, the assistant plugs deeply into system applications using Spotlight indexing, App Intents, and View Annotations to execute actions directly inside enterprise software, allowing users to manipulate live database objects like CRM records and invoices. Visually, the update introduces a redesigned UI orb that casts realistic environmental lighting onto nearby physical surfaces.

To process these heavy workloads on mobile hardware without triggering massive memory bottlenecks, Apple introduced its AFM 3 Core Advanced architecture to bypass standard DRAM limitations. Instead of forcing the entire 20-billion-parameter model to sit in active memory, the architecture stores the weights on the phone's slower NAND flash storage. When a prompt is entered, an initial routing step determines which specialized expert weights are required and loads only those specific components into DRAM, allowing active compute to scale dynamically between 1B and 4B parameters.

For complex reasoning tasks, a local System Orchestrator passes the requests up to Apple's Private Cloud Compute servers running on Nvidia GPUs within Google Cloud. Notably, while Apple used Google’s Gemini technology to train these models, it cut out third-party cloud runtimes entirely during execution to maintain a strict, privacy-first local stack.

On the developer side, Apple replaced Core ML with Core AI, a Swift-native framework built around AIModel and NDArray structures that compile directly into a proprietary .aimodel format.

This core architecture enables a suite of new system features, including plain-language script generation in Shortcuts, real-time explicit content blurring via computer vision, and an agentic Passwords app that autonomously navigates Safari to rewrite compromised credentials.

The Photos app also expands its local capabilities with generative tools like Image Playground, Clean Up, Extend, and Spatial Reframing, tracking all modifications using metadata and Google’s SynthID watermarking.

The downside comes down to steep hardware and regulatory limits. Running these models locally requires a baseline of 12GB of RAM, effectively locking the most advanced on-device features to newer hardware like the iPhone 17 Pro and iPhone Air.

Kimi Work runs up to 300 parallel local desktop agents to execute browser tasks and financial market workflows.

Voibe is private, offline dictation software for your Mac

Archyl uses AI to auto-generate interactive C4 architecture diagrams and drift metrics directly from your repository code.

Honen turns company documents, videos, and recordings into interactive, automatically updated AI employee training courses instantly.

ZeroGPU routes routine AI tasks to specialized small language models across a hybrid edge network for cheaper compute.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.