October 2026 Newsletter

Hot Off the Press

Anthropic (Claude)

The model lineup got refreshed. Opus 5.5 arrived September 22nd, with Anthropic saying it matches their Fable 5.1 model on most work while costing about 40 percent less to run than Opus 5. Sonnet 5.5 followed on September 29th, aimed at well-scoped jobs — one-pagers, summary slides, document edits, spreadsheet cleanup. Haiku 5.5 is now on for everyone on Team plans. I now find myself using Haiku for tasks that are procedural (“find this document for me”), Opus for almost everything else, and (rarely) Fable for very complex and/or nuanced analysis.

Claude can now work directly inside Google Docs, Sheets and Slides. This is the one I’d actually change my habits over. Instead of Claude handing you text to paste back, it edits in the file — revising a selection in Docs, building out a budget-versus-actuals sheet, cleaning up a deck against its own layouts. It’s on all paid plans and it’s a public beta. One catch for organizations: on Team and Enterprise plans an owner has to turn the connectors on before anyone can use it. If you’re on a free account, you don’t get this at all.

Anthropic added a rule against sustained cruelty toward Claude, taking effect November 12th. It explicitly doesn’t cover ordinary frustration, arguing with it, dark creative writing, or research. Anthropic is careful to say there’s no established evidence that models experience distress. I mostly find this interesting as a conversation starter in class, because of what it says about how people actually talk to AI. And I’ll repeat my advice about treating AI as a co-worker…I think it’s helpful to model the same behavior you’d use with a human, so saying please and thank you is an effective way to put yourself in the right frame of mind for effective interactions.

OpenAI (ChatGPT)

GPT-6’s “Intelligent UI” is rolling out to everyone, not just paid tiers. Responses now mix text with visuals and interactive elements rather than returning a bunch of text.

GPT-6 Sol and Luna are cheaper, faster siblings to Astra. Supposedly factual accuracy improved in these models, relative to Astra, but I haven’t tried confirming this yet.

OpenAI is retiring GPTs in December. If your org built a custom GPT — and a lot of nonprofits did, because for two years that was the obvious way to package “our AI helper for volunteers” — it goes away soon.

Google (Gemini)

Workspace Intelligence is on across Gmail, Drive, Chat, Docs, Sheets and Slides. You can ask it to summarize a folder, find a thread you can only vaguely describe, or draft a proposal from meeting notes plus a budget sheet. Google says it runs inside your own environment, only sees files you already have permission to see, and doesn’t train on your data. For a lot of nonprofits already living in Workspace, this is probably the single biggest change this month.

Docs and Drive now handle Markdown files natively. A small change, but it removes a common source of friction, since Markdown is the standard for AI-generated results.

Deep Review is a Google agent that gives rigorous critique on long technical documents. It was built for reviewing computer science papers and has expanded from there. It might prove useful for grant applications, evaluation reports or policy drafts, though keep in mind the academic-peer-review bias it was built with. Note that to access it, you go to https://research.google/deepreview/ and log in with your Google account.

Tip of the Month

Here’s a habit that’s changed how I work (thanks to my friend Susan for pointing me at Garry Tan’s talk on Personal AI).

At the end of a chat session, I stop and ask myself whether this is something I’m likely to do again. If the answer is yes, I don’t close the tab. I keep the same conversation going and say, “Let’s turn this into a skill.” Claude writes it up, and I can try it out immediately.

The reason this works better than sitting down to write a skill from scratch is that by the end of the session, I’ve already worked the process out. I’ve figured out how to talk to the AI about this particular thing, what steps it needs, where it goes wrong. All of that is sitting right there in the conversation. Writing a skill cold — before you’ve done the task even once — means encoding a process you haven’t actually tested.

This is a version of what Garry Tan, the president of Y Combinator, calls never doing one-off work. His rule is blunter than mine: “if you have to ask for something twice, you failed.”

Where his view and mine differ, is that I teach students to look at a task and ask whether it warrants the investment. There’s real time required to get a skill right, so is this something that happens often enough, or is big enough, to justify it? Tan would say stop making that decision. Don’t pick the tasks — just encode everything, because context pays off across almost everything you do. He’s not speaking hypothetically; he describes feeding in 25 years of emails, meetings, notes, drafts and the things he got wrong.

He’s also honest that it’s slow at first. He describes the first week as “you’re fixing more than you’re saving,” with the payoff showing up weeks later. That takes discipline, and I think it’s the more advanced move — one I’d like to come back to in a future newsletter.

In the meantime, I’m enjoying my growing library of skills that feel almost magical in how they are automatically invoked by Claude.

---

Follow the tip above far enough and you end up somewhere uncomfortable.

Suppose you do a genuinely good job of it. You build up the context an AI needs to work well for you — a pile of documents capturing your interactions and your experience. And you build up skills that capture your processes, the way you go about doing things. Tan describes writing these as Markdown files, and makes the point that this is really a way of programming an AI without writing code.

Now suppose you leave the organization, and the organization owns all of that.

They have something that is not you, obviously. But it is far more like you than anything an employer has ever been able to keep before. It isn’t just the documents you wrote or the programs you ran. It’s a working version of how you think about the job, which can be pointed at new problems after you’re gone.

I don’t think this is a reason not to do it. For a small nonprofit, where one person often carries institutional knowledge nobody else has, capturing that is arguably the responsible thing to do. But it’s worth going in with your eyes open, and worth a conversation that I suspect almost no employment agreement currently addresses.

Hans Peter Brøndmo writes about what he calls your digital soul, and the right to decide who gets a piece of it. His framing is that if your information is going to be shared, you should be the one choosing how much, with whom, and for what — so that “selling your soul” becomes a deliberate act you control, rather than something advertisers do to you quietly based on your online activity. He also takes up the harder question of what becomes of your digital soul after you die.

Both halves of that apply here. The context and skills you build are a more concentrated version of you than your browsing history ever was. There are going to be some very interesting conversations about who ends up with these pieces of your digital soul.