AI Product Updates Daily — June 8, 2026

WWDC 2026 opened today with the biggest rethink of Siri in the assistant's 15-year history. That was always going to happen. What made the day genuinely consequential is the surrounding layer of moves: Apple committed a billion dollars a year to Google's model stack, ended OpenAI's iPhone exclusivity, and — through a new Core AI framework — handed third-party developers on-device inference hooks that Core ML never offered. While Apple was on stage, xAI locked in an 18-month federal government deal for Grok 4, Anthropic's most capable (and most dangerous) model yet was quietly circulating among 150 critical-infrastructure organizations, and OpenAI updated the life-sciences fork of its model stack for the second time in a year.

Siri is now a separate app with a persistent chat interface, conversation history synced through iCloud, and access to personal data across mail, photos, files, and calendar.1 The underlying model is a custom 1.2-trillion-parameter system built on Google Gemini technology, for which Apple is paying approximately $1 billion annually. Dynamic Island integration ships on iPhone 16 and later; a system-wide "Search or Ask" gesture lets users invoke Siri without switching apps. The assistant can now read the screen and act across applications in a single request — the multi-step handoff between apps that the original Apple Intelligence promise described back in 2024 but never delivered.

The delay, which ran almost two years, came down to internal reliability issues Apple was unable to resolve using its own models. Handing the engine to Google is the resolution.

Apple is ending the single-vendor model for on-device AI. A new Apple Intelligence Extensions system lets users choose ChatGPT, Google Gemini, or Anthropic Claude as the powering model for any Apple Intelligence feature. Gemini is the default.2 This directly ends the exclusive integration OpenAI had from iOS 18 onward.

The competitive implication runs in two directions: Apple gains architectural flexibility, while Google, Anthropic, and OpenAI gain distribution across more than a billion active devices — a scale no API or app can match.

Apple is retiring Core ML in favor of a new Core AI framework that supports generative AI tasks and third-party model integrations natively.3 The current iOS 27 release opens text-only inference to third-party apps; image input capability is flagged for iOS 28. Apple also opened Siri extension APIs, giving developers a distribution channel into every device running iOS 27 forward.

Developer betas of iOS 27, macOS 27, iPadOS 27, watchOS 27, tvOS 27, and visionOS 27 are available today.1 iOS 27 includes a Liquid Glass transparency slider (addressing contrast complaints from last year's rollout), AI-powered photo editing with background extension and generative reframing, Safari tab organization by AI, and natural language Voice Control. macOS 27 drops Intel Mac support — Apple Silicon only going forward.

Tim Cook is delivering his final WWDC keynote as CEO. John Ternus — who built the M1 chip and the Neural Engine underpinning Apple's on-device inference — takes over as CEO on September 1. Cook moves to executive chairman.3 The iPhone Fold is expected to ship shortly after, making this transition one of the few CEO handoffs timed alongside both a platform reboot and a new hardware category.

xAI signed an 18-month OneGov agreement with the US General Services Administration, giving all federal agencies access to Grok 4 and Grok 4 Fast, plus dedicated xAI engineering support and agency training programs.1 The deal sits alongside a separate Pentagon CDAO contract with a $200 million ceiling that was announced earlier.

The practical outcome: xAI now holds a direct channel into all 18 major US federal agencies through at least March 2027.

Grok Build — the terminal-based AI coding agent for SuperGrok Heavy subscribers — remains in early beta but expanded its feature set this week.1 Key additions: parallel subagents for simultaneous multi-file editing, Git worktree support for isolated branch-level work, CI/CD pipeline headless mode, and ACP (Agent Communication Protocol) support for multi-agent orchestration. Install via curl -fsSL https://x.ai/cli/install.sh | bash. Grok Build competes directly with Claude Code (Anthropic) and Codex (OpenAI) in the agentic coding segment.

Project Glasswing — Anthropic's restricted membership program for critical-infrastructure organizations — now has more than 150 participating institutions, including several undisclosed Australian companies.4 Membership is limited to democratic-country organizations; selection criteria are not public.

The model at the center of the program is Mythos, which Anthropic describes as its most capable model to date in computer security tasks — effectively, offensive hacking. Mythos will not receive a public release. Glasswing members get access specifically to study defensive countermeasures before comparable models emerge from other labs.4

Anthropic's public blog post "When AI builds itself" — noting that Claude now authors over 80% of code merged into Anthropic's own codebase — also renewed calls for a global AI development slowdown, while acknowledging no viable coordination mechanism currently exists.

Anthropic also released the Claude Compliance API, integrating with enterprise security tools from CrowdStrike, Palo Alto Networks, and Okta. This is aimed at regulated industries where deploying an AI assistant requires audit trails and identity-layer controls the base API does not provide.

Released in early June and confirmed today, Lockdown Mode disables all of ChatGPT's network-dependent capabilities when activated: live web browsing, Deep Research mode, agent mode, file downloads, and most web-derived image generation.1 Core language model functions stay on. Personal users can toggle it under Settings > Security; enterprise admins can enforce it by role. The feature was designed for enterprise security teams managing data-leak risk in environments where GPT access is approved but internet connectivity isn't.

Dreaming V3 memory — the update that rolled out to free-tier users last week — shipped alongside Lockdown Mode.

GPT-Rosalind, OpenAI's life-sciences-specific model, received its second significant update on June 4, 2026.1 The update layers GPT-5.5's agentic coding and tool-use capabilities onto Rosalind's domain stack: medicinal chemistry, genomics, proteomics, spatial transcriptomics, applied genetics. On the GeneBench benchmark, the new version achieves 21.6% accuracy versus 20.4% for GPT-5.5 while using 31% fewer tokens. New additions include evidence retrieval and bioinformatics workflow plugins. Access remains restricted to vetted research organizations and the Rosalind Biodefense program — no public availability.

The AI bubble entered what analysts are calling its third inflation phase. Global equity indexes are up 20% from the March 30 bottom; the eight largest AI stocks are up 32%. Samsung, SK Hynix, and Micron each gained roughly six times over the past six months, driven by infrastructure demand.4 SpaceX has priced its IPO for June 12 at a valuation of approximately $1.75 trillion, with Anthropic's own S-1 in concurrent SEC review.5

The distinction analysts keep returning to — whether AI resembles the early internet (high current costs, transformative eventual returns) or a public utility (essential infrastructure, compressed margins) — has no answer yet. The compute infrastructure layer is already compressing toward utility pricing. The inference layer, where the big labs actually compete, has not.