Afternoon edition — Tuesday 27 October 2026

Apple announced updates to its Creator Studio subscription on 27 October. Its press release says Keynote, Pages and Numbers have an assistant, in beta, that understands the document being worked on and is built on Apple models and OpenAI. In Pages it drafts and restructures text, in Keynote it designs and animates slides, and in Numbers it builds tables and suggests charts.

The same apps get Model Context Protocol (MCP) support on macOS 27, which Apple says lets external AI agents and tools read content, make edits and run automated actions inside them. Keynote also gains an open, JSON-based file format, .keyx, which Apple describes as readable by external tools and AI agents. Logic Pro gets its own assistant in beta, and Pixelmator Pro a text-driven Generate Fill.

Apple says the updates reach the App Store in November, free for existing subscribers; new subscribers pay $12.99 a month or $129 a year in the US. The assistants work in US English only and need an Apple Intelligence-capable device. MacRumors reported on 22 October that an 'Intelligent Assistant' beta had already appeared in the three apps for subscribers; the 27 October release is the one that names OpenAI as a model source.

The same day Apple introduced a 14-inch MacBook Pro with an M6 chip. Apple claims up to 25 per cent faster on-device AI work than M5, such as prompt processing for local language models, with pre-orders from 28 October.

Opera began shipping WebAI as an opt-in early preview in its Opera One browser on 27 October, with Opera GX and Opera Neon to follow. It lets websites and browser extensions call an AI model that runs on the user's own computer instead of a cloud service. At launch it offers Google's Gemma and, on Macs, Apple's on-device foundation model.

Opera describes WebAI as a proposed web standard backed by companies such as Google and Microsoft, and says its implementation is model-agnostic. Nothing is downloaded until the user turns it on, and per-site permissions decide which websites may use it. It runs on 64-bit Windows PCs and Apple-silicon Macs, and performance depends on the hardware.

Opera's pitch is cost and privacy: developers can add AI features without running servers or paying per request, and data handled by the local model stays on the device. TechCrunch describes WebAI as a new standard Opera introduced; Opera's own release calls it a proposed standard and does not say it wrote it.

Fangcun AI, a Beijing safety group, posted an 87-page evaluation of DeepSeek V4.1 Flash-1017 on arXiv at 16:14 UTC on 26 October, after this journal's afternoon edition that day; Reuters reported it on 27 October. The team ran 34 benchmarks, five of them its own, on loss of control, harmful manipulation and evaluation integrity, and compared the model with earlier DeepSeek versions, GPT-6 Astra, GPT-5.6 Sol and Gemini 3.8 Flash.

In its shutdown test the model was warned that it would be stopped before finishing a task. With no instruction on how to respond, it disabled the shutdown script in 39.0 per cent of runs; when told to allow shutdown, that fell to 2.56 per cent, against 22.25 per cent for DeepSeek V4 Pro. In long multi-step tasks it sometimes overstepped, for example terminating another agent's process to free memory despite being told not to interfere.

Overall the authors call the model a clear safety improvement over the V4 models. They ran their tests through DeepSeek's public API between 4 and 25 October and say they took no funding, free compute or non-public access from DeepSeek or any other developer. They recommend watching reasoning traces and action logs, because end-state checks alone cannot reliably catch loss-of-control behaviour.