AI News — June 9, 2026

Model releases

• [2026-06-08] Apple / Google — the rebuilt Siri AI shipped at the WWDC keynote running on the custom ~1.2-trillion-parameter Gemini model, behind a three-tier routing stack: simple requests stay on-device, moderately complex ones go to Apple’s Private Cloud Compute, and the heaviest reasoning routes to Google Cloud (on Nvidia B200 GPUs), with queries anonymized at each hop so neither Apple nor Google can tie them to a user. Siri AI is also a standalone, more conversational app with personal-context and onscreen awareness across iPhone, iPad, Mac, Watch, and Vision Pro. It ships with iOS/iPadOS/macOS 27 this fall, but not in the EU or China at launch while Apple works through regulators. (official, source)

  This is the year’s largest single distribution event for a frontier model — a custom Gemini reaching a billion-device consumer surface — and an Apple admission that its own foundation models weren’t ready to run the assistant. Routing through Private Cloud Compute so Google receives only anonymized, un-trainable traffic sets the template others will be measured against when they want a rival’s model on a sensitive surface without handing over the data.

• [2026-06-08] Apple — the genuinely new piece beyond the pre-keynote reporting is Apple Intelligence Extensions: a framework plus App Store marketplace that lets users set Claude, ChatGPT, Gemini, or Grok as the model backing Siri and system AI features (Writing Tools, Image Playground), each replying in a distinct voice so you can hear which model answered. This confirms yesterday’s unconfirmed model-choice report — and adds Grok and the marketplace mechanics — while ending Apple’s single-provider ChatGPT arrangement in favor of an open, swappable backend. Developers wire their model in via App Intents and ship it as an installable Extension; it arrives with the fall 27 releases. (official, source)

  Recast as a marketplace, “the default model on a billion devices” becomes a contested slot the way the browser’s default search engine is — direct distribution that Anthropic, OpenAI, Google, and xAI now compete for inside Settings. For builders, the integration contract is the thing to watch: an App Intents-based Extension is a concrete, Apple-blessed surface to target, not just an API key.

AI cost tracking & telemetry

• [2026-06-08] FinOps Foundation / Tokenomics Foundation — at FinOps X (San Diego, June 8–11), last week’s “intent to launch” turned into something teams can adopt: FOCUS 1.4 — the open billing spec for normalizing cost data — was released at the conference, and the roadmap names v1.5 as the version that will break AI spend down by token type and workload, so inference cost can be tied back to the team consuming it. The newly formed Tokenomics Foundation (Linux Foundation, in partnership with the FinOps Foundation) detailed its initial working groups and technical roadmap on stage. (source, source)

  The practical promise is a common schema for token spend: today every provider and gateway reports usage its own way, so cross-vendor attribution is bespoke glue code. A FOCUS extension that treats tokens like any other billable unit lets token spend flow into the same dashboards and showback as cloud cost. The caveat is timing — 1.4 is here now, but the AI-specific token breakdown is v1.5 roadmap, not shipped.

Watch list

• Gemini 3.5 Pro (this month): still unshipped — promised “next month” at I/O (May 19), reportedly a 2M-token context window, a “Deep Think” reasoning mode, and pricing around ~$15/$60 per 1M tokens, with no model card, benchmarks, or API name yet. Watch all four on day one. (source) (unconfirmed)

• Apple Intelligence Extensions (fall 2026): with the marketplace confirmed, the open questions are which models actually ship as Extensions at GA, what the App Intents-based developer API looks like in the iOS 27 SDK, and whether the EU and China get it later in the cycle. (source)

• MiniMax M3 weights (~June 11): M3 launched June 1 with frontier coding claims (59.0% SWE-Bench Pro) but no weights or technical report; MiniMax committed to publishing both on Hugging Face/GitHub “within 10 days,” so the real open-weights test and independent benchmarks land around June 11. (source) (unconfirmed)

• xAI Grok V9-Medium (mid-June): xAI reportedly finished training a ~1.5-trillion-parameter model — ~3× its current production model, trained on Cursor developer-workflow data and aimed at the coding lead — with a public release expected mid-June. Now extra relevant as one of Apple’s four named Extension options. No model card or independent benchmarks yet. (source) (unconfirmed)

• Anthropic developer billing split (June 15): Claude Agent SDK, claude -p, Claude Code GitHub Actions, and third-party agents reportedly move off subscription limits onto a separate monthly credit pool (~$20 Pro / $100 Max 5x / $200 Max 20x), metered at full API rates with no rollover; Anthropic still hasn’t detailed it officially. The sharpest impact is on bursty automation — per-commit CI and GitHub Actions — where full-rate metering turns a usage spike into a bill spike. (source) (unconfirmed)

• MCP spec 2026-07-28 final: stateless core, Tasks, and MCP Apps move to stable, and clients must validate the iss parameter per RFC 9207. Server operators relying on sticky sessions will need to migrate — worth auditing deployments (and iss handling on clients) before the date, especially after the recent VIPER-MCP server-side CVE wave. (source)