Claude Ships an Official Package for Apple’s Foundation Models Framework
Anthropic released anthropics/ClaudeForFoundationModels, an Apache-2.0 Swift package that conforms Claude to Apple’s Foundation Models framework on iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27. You drive it with the same LanguageModelSession API used for Apple’s on-device model — respond(to:), streaming, guided generation, and tool calling all behave identically.
The practical payoff is provider portability: a team can prototype on Apple’s on-device model, then route heavier queries to Claude (or Gemini) by changing a Swift Package Manager dependency, with no edits to session logic. Server tools like web search, web fetch, and code execution run on Anthropic’s infrastructure in a single round trip. It’s explicitly beta and tracks the OS 27 server-side API, so expect changes before GA.
Enterprise Custom Roles Can Now Delegate Admin — Without Handing Over the Org
Anthropic added admin permissions to custom roles on Enterprise plans. Owners can now grant a role scoped access to specific admin areas — Identity & Access, Billing, Analytics, Privacy, User Management, and Libraries — at three levels: No access, Can view (read-only), or higher. So a finance lead can see billing without holding the keys to everything else.
It’s opt-in by design: if you don’t configure admin permissions on a role, it grants no admin access and administration stays with Owners and Primary Owners. Unglamorous, but exactly the least-privilege delegation that large, audited organizations need before they widen Claude access across teams.
New Tool Versions Trim Tokens and Expose the Code-Execution Limit
The Claude Developer Platform shipped fresh tool versions for code execution, web search, and web fetch. code_execution_20260521 now surfaces the 90-second per-cell execution limit directly in the tool description, so Claude can budget long-running cells instead of getting cut off mid-task.
The bigger win for agent builders is response_inclusion, added to web_search_20260318 and web_fetch_20260318. It lets you drop already-consumed result blocks from the API response — no beta header required — which keeps context windows lean across multi-step agentic loops. Small changes, but exactly the kind of plumbing that decides whether a long-running agent stays inside budget.
Session Polish: Conversation-Language Titles and a Cleaner Remote Control
Claude Code’s latest release sweats the small stuff. Session titles now generate in the language of your conversation (pin one with the language setting), and Bedrock credential caching now holds credentials until their actual expiration instead of a fixed one-hour window — fewer surprise re-auths mid-session.
Remote Control got the most fixes: connecting from web or mobile no longer silently switches the session’s model, disconnect notifications show a human-readable reason instead of a bare numeric code, and signing into a different account now properly disconnects the session. None of it is headline-grabbing — all of it is the friction that daily Claude Code users actually feel.
Fable 5 and Mythos 5 Stay Offline on Day 9 — and the NSA Detail Changes the Story
Anthropic’s two frontier models remain suspended worldwide, now nine days into the US export-control directive issued June 12. Anthropic’s International MD Chris Ciauri says the company is “very confident” the models return “in the coming days,” but they’re still dark. Prediction markets price roughly 57% odds of restoration before July 1.
A new detail reframes the dispute. Anthropic originally characterized the trigger as a narrow jailbreak it disagreed warranted a recall — but reporting now says the NSA, in a red-team exercise, watched Mythos autonomously breach nearly all of its classified systems within hours. If accurate, the suspension is about a model’s offensive cyber capability as a whole, not one exploit — a far harder thing to patch around.
Claude Code and Cowork Can Now Drive Your Desktop
Anthropic expanded Claude Code and Claude Cowork with a computer use capability that lets Claude directly control a Mac or Windows desktop — clicking, typing, opening apps, and navigating browsers to finish UI-heavy workflows. It’s a research preview for Pro and Max subscribers, and notably it doesn’t reach for the mouse first: it prefers existing connectors like Slack or Google Calendar, steps up to browser control next, and only takes direct desktop control as a last resort.
Pair it with Dispatch — a companion that lets you assign tasks from your phone — and the shape of “always-on assistant” comes into focus: check email every morning, pull weekly metrics, or spin up a Cowork or Code session for a report or PR without sitting at a terminal. For developers who already live in Claude Code, it means delegating the legacy-app and browser drudgery that connectors don’t cover.
Apple Just Made the Model a Swappable Part — and That Cuts Both Ways
By conforming Claude to Apple’s LanguageModelSession protocol, Anthropic gets distribution into every iOS 27 developer’s toolbox — a team can reach for Claude with one Swift Package Manager line and no rewrite. That’s an enormous top-of-funnel win, the kind of default placement money usually can’t buy.
But the same portability that gets Claude in the door makes it trivial to show it out. When Apple’s on-device model, Gemini, and Claude are interchangeable behind one API, the model becomes a commodity slot, and the contest moves to capability per dollar, latency, and the surrounding platform — not lock-in. Anthropic is betting it wins that bake-off on quality. It’s a confident bet, but it’s a bet the abstraction layer forces it to keep making, app by app.
When Frontier Models Become Munitions, Global Ambition Gets a Ceiling
Nine days into the Fable and Mythos freeze — now with an NSA red-team detail attached — the signal is hard to ignore: the most capable models may be export-controlled by default. For a company eyeing an IPO and aggressive international growth, that’s a structural risk, not a one-off. It implies a tiered product line — full capability at home, throttled capability abroad — with national-security policy acting as a gating function on the roadmap itself.
Last week the risk factors were reliability and billing. This week it’s sovereignty. A frontier lab’s flagship being treated like a controlled munition reframes every overseas deal: customers want the best model, and Washington increasingly gets a vote on whether they can have it. Expect “model availability by jurisdiction” to become a standard line in enterprise contracts — and a standard question from investors.