The Signal: AI updates for September 21, 2026

Claude Docs and Slides, ChatGPT in Word, Gemini Live and a reasoning-setting change: what changed, who gets it, and useful things to try.

September 21, 2026 · Better ways into everyday work

This week, the interesting question is where you meet AI. Inside a document? In a conversation you can interrupt? Alongside a project you return to tomorrow? Here are four changes worth understanding and one development in how AI is evaluated. This is a review of selected primary announcements available on September 21, not a complete news feed or an independent product test.

Claude brings documents and slides into the conversation

Anthropic announced a merger of Claude chat and Cowork on September 16, beginning with a staged Pro and Max rollout. Docs and Slides also launched in beta on paid plans; Enterprise administrators control activation. An announcement date is not a promise that your account already has the feature.

The opportunity is to carry one set of facts into several useful formats. Our new walkthrough uses a fictional repair café: a short brief becomes a document and a five-slide introduction. The room has 24 seats, only 18 places are bookable, and the venue is still unconfirmed. Those details make it possible to check whether an attractive deck stays faithful to its source.

Try the document before the slides. Make one factual correction, then check whether it survives the next format. Treat a changed date or an invented booking link as a failed check, even when the layout looks excellent.

Sources: Anthropic · Claude Cowork and chat are now one Claude, September 16, 2026

Try the document-to-slides walkthrough

ChatGPT arrives in Word

OpenAI’s September 17 notes introduce ChatGPT for Word through its Microsoft add-in, including Free-plan access subject to usage limits. It can help draft, summarize and revise within Word.

If your document already lives in Word, a useful first test is a single paragraph. Ask for a clearer version that preserves every number, name and qualification. Compare the result with the original before changing the rest. A request to improve readability should not quietly become permission to strengthen a claim.

Keep an untouched copy of a document you care about. After a larger revision, inspect headings, tables, captions and the exported file as well as the prose. The place where you edit and the file a reader receives can expose different problems.

Sources: OpenAI · ChatGPT release notes, September 14–17 entries; checked September 21, 2026

Build a document review habit

Check your reasoning setting for harder work

The September 14 ChatGPT notes say automatic switching from Instant to Thinking is being retired for Plus and Pro users. Available reasoning options can still be selected in the model picker; automatic safety switching is a separate exception.

When a task involves reconciling conflicting evidence or checking several constraints, look at the selected mode rather than assuming that a complicated question will change it for you. More deliberation is still not verification. Ask for the evidence, calculation or finished artifact you can inspect.

A useful personal comparison is one ordinary question and one harder task with a known answer. Keep the inputs the same, record the mode you chose, and judge whether the result helped. Do not turn one good response into a universal model ranking.

Sources: OpenAI · ChatGPT release notes, September 14–17 entries; checked September 21, 2026

Try a small, repeatable comparison

Gemini’s new voice models invite a different kind of practice

Google announced Gemini 3.8 Live and Live Extended Thinking on September 15 and updated the announcement September 17. The rollout distinguishes Search Live, Gemini Live, Workspace subscription access and developer APIs. Check the relevant app and account rather than treating every surface as interchangeable.

Our suggested experiment is a five-minute spoken explanation. Describe a hobby to a curious beginner, ask for one follow-up, then correct a detail halfway through. At the end, request a short written recap. Did it retain the correction? Did you get to finish your thought? Was the recap faithful to what you actually said?

A friendly voice can make a session enjoyable. The useful measure here is whether you explained something more clearly and can repeat that explanation without the assistant.

Sources: Google · Gemini 3.8 Live and Extended Thinking, September 15; updated September 17, 2026

Run the five-minute voice exercise

Read safety announcements for what is still unresolved

On September 18, Anthropic announced an embedded-evaluation partnership with Accenture, led by Faculty. Anthropic says it will fund this work directly and that access and reporting standards are not yet settled. This is an announced evaluation arrangement, not a completed independent certification of model safety.

When you read an AI safety claim, ask who performed the evaluation, who paid for it, what they could inspect and what findings will be public. Those questions help you distinguish a commitment from a result. For your own work, retain a clear review point before an assistant sends, publishes or changes something consequential.

Sources: Anthropic · Accenture embedded evaluation, September 18, 2026

Set a useful action boundary

Take this into your next task

Choose one change that fits something you already do. Try a small task, inspect the result, and keep the method only if it helps. You do not need to adopt every new feature to become better at using AI.

Build a project brief.

Questions people ask

Is this advice based on hands-on testing of these releases?

No. Release facts are attributed to the linked providers. The exercises are original teaching suggestions with fictional inputs, not recorded benchmark results.

Will these features be available in Canada?

Availability can depend on the product surface, rollout, plan and workspace controls. Check the linked announcement and the options visible in your account; this edition does not assume that a US-only feature is available in Canada.

Sources and editorial notes

Reviewed 2026-09-21. Product capabilities depend on the app, plan, region and workspace settings. Workflows and prompts are authored teaching material, not recorded model results or measured time-saving claims.

Keep going

Related Power of AI pages

Keep reading with Start here, everyday uses, the tool guide, the writing workshop, and sources and standards.