Apple Intelligence is cleared for China — test build lands this month
Apple's AI stack finally has a mainland launch date, Anthropic put a watchdog inside its own building, and Perplexity handed its agent sandbox to everyone else.
Apple executive chairman Tim Cook said Apple Intelligence has been approved for the Chinese market, with a test version landing later this month — and Siri AI is not in it. Speaking in Shanghai on Thursday, Cook named the first batch of features as real-time translation, writing tools and the photo Clean Up tool, and said the approval is what lets "the first core Apple Intelligence features slowly come online." Bloomberg reported Wednesday that the full rollout arrives in December through the iOS 27.3 update, shipping the feature set Apple debuted in the US two years ago rather than this year's redesigned assistant. The regulatory path is now public: a generative-AI filing on July 8, then network-access permits from the industry ministry on September 29 naming three approved systems — Apple Intelligence alongside Alibaba's Tongyi Qianwen and Tongyi Wanxiang — covering the iPhone Duo, the iPhone 18 Pro and earlier models. Baidu carries the search and Visual Intelligence work, per the same filings and Bloomberg.
Why it matters: Apple is entering the most competitive AI market on earth with a two-year-old feature set and no assistant, while Huawei, Xiaomi, Oppo and Vivo have been shipping on-device AI to those same buyers for a year. Cook gets a launch; he does not get parity. The interesting question is what ships in the second Chinese update — because that is where Siri AI either arrives or becomes the first Apple product permanently withheld from a major market. We flagged the December timing as a report yesterday — Today in AI — October 14.
Anthropic has onboarded six METR researchers with company laptops, office space and internal tools roughly on par with what its own risk-assessment teams use — and given up the right to block what they publish. The arrangement, announced in an Anthropic blog post on Thursday, makes METR the first evaluator at the top of a three-tier access system: tier one runs narrow reviews and needs Anthropic's approval to publish (biosecurity nonprofit SecureBio sits there now), tier two can publish without approval, and tier three adds the employee-level access. Anthropic can require redactions only in pre-defined categories — customer data, attorney-client privileged material, model architectures — and METR may publicly say whether it thinks each redaction was justified.
This is the promise Dario Amodei made in a September essay becoming an operating arrangement, which is the harder test: the terms are now written down and reviewable, and OpenAI, Google, Meta and Nvidia have all signed a White House accord pledging some version of independent review. The structural gap we flagged in September is unchanged — evaluators can investigate and report, not stop a release (The AI watchdog gets a badge, a laptop — and no veto power). The new detail worth watching is tier two: if a mid-tier evaluator publishes something unflattering without approval, we learn whether Anthropic's redaction categories are narrow or elastic.
Perplexity open sourced SPACE, the sandbox its own agents have been running in since July, and signed CoreWeave, Akamai and Crusoe to offer it to their customers. SPACE — Sandboxed Platform for Agentic Code Execution — puts each agent's code inside a Firecracker microVM with its own kernel, so a runaway agent hits a virtual-machine boundary rather than the host, and credentials and network access are managed outside the sandbox so nothing sensitive lives in the working environment. It is written in Rust, speaks HTTP, ships an E2B-compatible Python wrapper, and carries the Apache 2.0 license. Perplexity says it has powered every Perplexity Computer session since July at millions of sandbox creations a day, that creation runs three to five times faster than with its previous commercial provider, and that 108 escape attempts in its own evaluation never got from the virtual machine to the host.
Read the escape claim with the usual caution — Perplexity testing Perplexity — but the distribution is the news: an agent-sandbox vendor just made its isolation layer a commodity across three clouds, which is the cheapest way to make every competitor's sandbox look under-tested.
What to watch: whether any tier-two Anthropic evaluator publishes without approval, and whether the December China build actually ships on schedule.
Apple is entering China's AI market two years behind its domestic rivals. Does a late launch still count as a launch? Tell us in the comments.
