Hello fellow keepers of numbers,

I’m writing this a day early this week because I’m traveling all day Friday. I’m one of those fantasy football crazies, and I travel back to my hometown every year for an in-person fantasy football draft with some of my childhood and high school friends. If you don’t understand fantasy football and aren’t a die-hard sports fan, this probably sounds like the most ridiculous trip of all time. Which is exactly why my wife thinks it’s the most ridiculous trip of all time.

This week, we were introduced to the Last Accounting Company. Which is obviously a lie. But they’re at least executing something I find mildly interesting. We also finally (!!) got a ChatGPT Business Premium tier. For those who’ve been loving the new ChatGPT app, this is a huge deal.

Plus, stick around to see how to set up a new event-driven scheduled task in ChatGPT. For example, trigger an automation when you receive a new email.

The Latest

Last Accounting Company launched an agent-native accounting firm that combines accounting services and software in one offering. The company says AI agents perform the underlying work while human accountants validate the results before anything is filed.

The agents keep the general ledger continuously updated, handle bookkeeping, invoices, payroll, and receipts, and prepare filings according to local requirements. The system connects directly to banks, payroll platforms, and internal tools, and customers can query live financial information through Slack or MCP.

Unlike an accounting software provider, Last Accounting Company performs the accounting work itself and operates its own general ledger and agent stack. Y Combinator lists the Helsinki-based company as active and part of its Summer 2026 batch. The service is targeting startups and SMBs in Europe and the U.S., and pricing has not been published.

Why it’s important for us:

I like the idea behind this a lot more than I like the name. Last Accounting Company is a terrible name for obvious reasons.

This is one possible vision for the future of accounting firms: bring a client’s financial data into one place, use agents to categorize transactions and handle repetitive work, and give the client real-time access through messaging tools, MCP, dashboards, forecasts, cash-flow models, and other tools. Most firms aren’t going to build their own ledger anytime soon, but they can still move toward that vision by connecting their existing systems and automating pieces of the work.

But getting the data together is often very difficult. Real businesses don’t have pristine systems with APIs that all happily talk to one another. Their data is scattered across old software, spreadsheets, and inboxes. There are often several manual steps, and that usually lives inside the brain of one or two people. If connecting it all were easy, accountants would’ve solved this problem a long time ago.

We're seeing more news stories like this one every week now. They all intentionally ignore the complexity because that doesn't make the business sound sexy. Agents might be capable of doing some of the accounting once the data arrives, but good luck convincing the prospective client they should go through 5 major software implementations to upgrade their systems just so you can fully automate the accounting that they can have done for the same price by the CPA who offices down the street from them who'll start on that this month.

I don't think the direction is bad, but the approach feels wrong. I think I'm pretty confident this won't be the last accounting company.

OpenAI launched Premium seats for ChatGPT Business on August 25. Premium seats include 5x more usage than Standard seats and remove the five-hour usage limit.

Workspace owners can mix Standard and Premium seats in the same workspace, assign or reassign them as usage needs change, and monitor usage and spending centrally. Both seat types include access to ChatGPT and Codex, and workspace credits can extend usage beyond the included limits.

Premium seats cost $125 per user per month or $100 per user per month when billed annually. Standard seats remain $25 monthly or $20 per month when billed annually. ChatGPT Business still requires at least two paid seats, but those seats can be any combination of Standard and Premium.

Why it’s important for us:

Just got off the phone with Sammy Altman, and he said I complained about this enough that they finally launched a ChatGPT Business Premium plan.

In reality, I think this was very heavily requested by most power users in the business world. Until now, ChatGPT Business only had a $20 per month plan. Once you hit the limit, you paid for additional credits at API rates. That could get expensive fast.

Claude has the Team Premium plan that costs $100 per month. I've used that for as long as I've had Claude. I occasionally hit limits, but not often enough that it infuriates me like it did with ChatGPT.

Based on estimates I ran a few months ago, the same work that fit within Claude’s $100 plan would have cost me roughly $400 to $600 each month in ChatGPT. That gap was big enough to make Claude the obvious choice for anyone that was a power user.

This new premium option changes that. ChatGPT is now extremely viable for accounting firms with heavier users, and I think its desktop app has surpassed Claude’s overall. Claude still has a slight edge working with Microsoft Office apps, some accounting tasks, and the simplicity of Cowork. But the gap isn’t very big anymore.

There's no going wrong with either option, which is exciting to have some newfound competition in the enterprise space.

Trending News

Thomson Reuters launched Thomson, its proprietary frontier model trained for legal and tax work: They've specifically noted that it's cheaper than most frontier models, which is certainly a positive. But it'd just be so much better if they opened up their tool to be accessible by Claude, ChatGPT, and others. That won't happen soon, if ever, though.

Xero announced an OpenAI connector and plans to expand live Xero data into Excel, Word, PowerPoint, and Copilot Cowork: I'd love for this connector to be awesome and all of this work flawlessly. After the previous failure (can we call it that?) of JAX, and maybe some PTSD about how awful the QBO connector is, I'm skeptical until I see it in action.

OpenAI enabled secure website sign-in for ChatGPT Work’s cloud browser: AI controlling the browser is inevitable. There are still security concerns here. Hopefully they'll be alleviated soon because it unlocks so many amazing possibilities. Even providing a login right now feels risky though because then it has access to all data within that account.

OpenAI added event-triggered scheduled tasks that can respond to Gmail, Slack, and GitHub activity: This is an extremely exciting step forward for scheduled tasks. Until now, they've been time based. But think of a world where you can create scheduled tasks that automatically run when you receive an email, or when a form is submitted, or when you receive a Teams message.

Anthropic unified Claude’s memory across Chat and Cowork: If you use memories in Claude, this is a nice quality of life improvement.

Nvidia reportedly agreed to acquire Hugging Face for $12.9 billion: Hugging Face houses a large number of AI models, many of them open source. I think Nvidia sees OpenAI, Anthropic, and Google all investing in both AI models and chips, and maybe felt a little bit of FOMO? This is an interesting acquisition nonetheless.

OpenAI cut GPT-5.6 Sol pricing by more than 20% for three months: This is yet another indicator that AI is improving itself. OpenAI recently cut Terra and Luna prices because of efficiencies gained using their models to improve itself. Now it's the same for Sol.

Z AI revealed that the anonymous Ox Alpha model was GLM-5.3-Flash, a low-cost open model run on Chinese chips: This model is supposedly really good and also really cheap. It's noteworthy because it's using Chinese chips and they have a lot of capacity.

Sam Altman said OpenAI expects to have something it internally calls AGI by December: I didn't know we were still on AGI watch. Feels like this is a cheeky marketing opportunity pre-IPO. But maybe there's some major new leap we should be expecting in the next 4 months?

Put It to Work

Because of my travels, no video this week. But I have a simple demo of event-driven scheduled tasks in ChatGPT.

The new scheduled tasks can run on a specific trigger if you use the Gmail, Slack, or GitHub plugins. Right now, these only work on web and mobile. I hope, and expect, this will be available in the ChatGPT desktop app soon.

Here’s an example of how to create an automation that can trigger when you receive a new email. In this case, I’ve told it to identify if it’s coming from a specific person because I’m expecting something urgent.

I prompted it to trigger when I receive an email, check who it’s from to filter for the appropriate contact, and then notify my team with the contents of the email.

It created the scheduled task for me in 27 seconds. Now, time to test it…

Here’s the fictional email:

There was about a 1-2 minute delay, but it ran successfully.

Overall, it took me less than 5 minutes to complete end-to-end.

Weekly Random

Amazon is buying old and rare books, scanning them to train AI, and destroying the physical copies in the process.

Reporters tracked one of the books from a roughly 1,000-book order to an Amazon warehouse in Las Vegas. Workers there reportedly cut off the bindings so the loose pages can move through industrial scanners more quickly. The original books are ruined.

This is disgusting...

Obviously, Amazon has the right to buy books legally. But choosing to extract the data and destroy the work so that nobody else can use it to train AI models is so unethical that it's hard to put into words. And Amazon likely isn't the only one doing this.

Some of these books are rare or out of print, and there may only be a few surviving copies left. This type of stuff is one of the reasons why AI has such a negative stigma.

AI-generated content is already everywhere. It’s taking over social media, photos, videos, music, and even books. It’s not unreasonable to imagine a world a few years from now where most of what we see online was created by AI, then recycled into more AI-generated content. The dead internet theory doesn’t sound quite as ridiculous anymore.

And if that world is coming, the human work left behind becomes even more important. These books are physical pieces of what people researched, imagined, wrote, and created before machines started filling every screen with unlimited content.

We’re feeding that human history into AI models and then throwing the originals in the trash.

It makes me sad. And I hate it.