New AI tools: what launched in late October 2026

The new AI tools of late October 2026 are mostly about putting assistants and agents inside software people already use: a Git forge built for coding agents, an assistant in Apple's office apps, and a world model from Google DeepMind. Here is what launched, who it is for and what is still unknown, with a link to our full report on each.

We track launches in the AI tools section and work from the vendors' own announcements. Everything below was checked against those pages in late October 2026.

New AI tools at a glance

Entire: a Git forge for agent-human teams

Entire, led by former GitHub CEO Thomas Dohmke, opened its forge on October 26. Its announcement describes three parts. Trails take the place of the diff-based pull request, with evals that agents score on each push. The hosting layer is a Git-compatible server that can replicate a repository across regions. Continuous Memory syncs agent sessions and prompts into the repository so later agents can search them. Buildkite and Depot are the first CI partners. Entire's speed claims come from its own benchmarks. Our report: Entire launches a Git forge for AI agents and developers.

Apple's assistant in Keynote, Pages and Numbers

Apple announced on October 27 a beta assistant in its three office apps, built on Apple models and OpenAI. It reads the open document: it helps design and animate slides in Keynote, draft and restructure text in Pages, and organize and chart data in Numbers. The apps also gain Model Context Protocol support, so outside AI agents can work with documents on macOS 27 or later. Apple lists the assistant as a feature for Creator Studio subscribers, in U.S. English only. Read the full story.

Genie 4 from Google DeepMind

Genie 4 is a world model: it generates a scene you can steer while it runs. DeepMind says it produces music, ambient noise and dialogue together with the visuals, and it builds on models from the Gemini family. Access starts with a waitlist. Our Genie 4 report covers what DeepMind says it has not solved yet.

Unlimited OCR 2 from Baidu

Baidu published Unlimited OCR 2 under the MIT license, with weights on Hugging Face and code on GitHub. It is built to transcribe a long document in a single pass rather than page by page, which suits contracts and scanned books. Baidu's benchmark claims have not been tested independently. Details and how to run it.

Gemini Notebook, the renamed NotebookLM

Not every new name is a new product. Google renamed NotebookLM to Gemini Notebook and calls it the same standalone research tool; notebooks and shared links carry over. Google's plans page still lists it on the no-cost tier. What changed and what did not.

How to read a launch

A launch post is a claim, not a test. For each new tool we look for three things: who can use it today, what the vendor admits is missing, and which numbers come only from the vendor. If a tool is in beta or behind a waitlist, the honest answer to "should I switch" is usually "not yet". When a vendor's own changelog is enough for you, read that directly; when you want the context in one place, the ClaudeGPT Si RSS feed collects each story with its sources listed at the end.