Hello fellow keepers of numbers,

OpenAI hosted its DevDay 2026 this week. I’ve done my best to cover the most interesting things they announced. We also got the new Sonnet 5.5, which includes some of the same improvements we’ve seen from Opus 5.5. Lastly, Numeric launched their NetSuite competitor. It’s definitely difficult to build an ERP, right? Look at NetSuite. They’ve been trying for 3 decades, and it still doesn’t work.

Plus, stick around for a demo showing how to use Claude Docs, Claude Slides, and Claude Design.

The Latest

OpenAI announced a slate of new features at its September 29 DevDay, with another Sites release following the next day.

Why it’s important for us:

OpenAI's DevDay included lots of awesome announcements. I'll break down each of them quickly:

(1) GPT 6.1 Sol

I’ve been using this quite a bit over the last couple of days, and the efficiency improvement is great. I still don’t think it’s as good as Opus 5.5, but it consumes very little usage.

Sonnet 5.5 might be the closer comparison. If Sol is smarter and consumes close to the same usage, that’s interesting. Either way, I really like what they’ve done here.

(2) Dots and specialist dots

These are basically OpenAI’s equivalent of Grok Bot or Meta’s Muse. It's an agent that’s just always there.

Honestly, I don’t really understand how this is different from ChatGPT Work, which already has its own cloud computer. Maybe it has a personality and that's what makes it interesting? I don't know. It doesn't immediately appeal to me.

Other people could find this very interesting for the same reasons they’re loving Grok Bot and Muse. The specialist dots are more interesting to me. If OpenAI deploys prebuilt agents for specific business tasks, I like that idea a lot.

(3) Space and Pages

This is essentially the same thing Anthropic released recently with Claude Docs and Claude Slides. I love the idea of my documents living inside my AI tool. Working alongside your AI in the same document is so much easier than going back and forth with a Google Doc, Google Slide, Word Doc, or Excel file.

I think this is how I’ll work most of the time. Accounting firms have more considerations around client deliverables and internal documentation, so maybe it’s not quite there for firms yet.

But if this continues to improve, I think it could be a big disruption in the document management space and the files we typically work in.

(4) Sites

I’ve said this was coming a few times over the last couple of weeks, and it’s finally here: ChatGPT Sites with plugins.

The ability to connect a site to tools like Outlook or SharePoint is what makes this so interesting to me. It’s the same reason connecting Claude artifacts to those tools made them much more useful. This is a huge update. It makes Sites insanely powerful.

(5) ChatGPT in Slack and Teams

The idea of having ChatGPT available in Teams to @mention is obviously enticing. But I thought tagging Claude in Slack was going to be a pretty big deal, and I really haven’t ever used it.

I’m wondering if this will be the same for me. For people who like working that way, though, having it available in Teams is nice.

(6) Sign in with ChatGPT

This lets you use your ChatGPT usage inside another tool’s AI features, assuming that tool supports it. For example, you can use your ChatGPT usage for AI features in Notion.

(7) MCP events

I think this is probably the most underrated announcement of the entire DevDay. MCPs are what let ChatGPT connect to other tools through plugins. For example, using ChatGPT to read your Outlook. Events let you tell ChatGPT what changes to watch for and what to do when they happen.

If you're familiar, think of Zapier or n8n. Changing a tracker’s status from “not started” to “active” could trigger ChatGPT to kick off another workflow. That’s now possible with tools that support MCP events.

The tools will need to build support for those events into their MCPs, so it won’t automatically work everywhere. But I think this is a huge next step in automation with AI.

Anthropic launched Claude Sonnet 5.5 on September 28, the second model in its Claude 5.5 family. The company says it generates output more than 30% faster than Sonnet 5 and costs up to 30% less per task in its testing.

Sonnet 5.5 improves document, spreadsheet, and presentation work. In one internal test, Anthropic gave it a public company’s quarterly earnings materials, call transcripts, and a slide template to create a 10-slide operating review. Two experts judged the first draft ready to send without changes.

Anthropic says Sonnet 5.5 approaches Opus 5.5 on several evaluations of professional work, though Opus remains stronger on complex, open-ended tasks requiring sustained judgment. The company positions Sonnet as a faster, lower-cost option for clearly defined everyday work.

API pricing remains $2 per million input tokens and $10 per million output tokens. The reported savings come from using fewer tokens to complete tasks. Sonnet 5.5 is available now across Claude’s products, its API, Amazon Web Services, Google Cloud, and Microsoft Azure.

Why it’s important for us:

After last week's Opus 5.5 release, I was hoping Sonnet would get some of the same improvements. And it did.

I've only done some early testing, but it already feels like a pretty big jump from Sonnet 5. It seems smarter, and the efficiency improvements are noticeable for me. It should help get a lot more work done before hitting usage limits.

What interests me most is Anthropic specifically calling out improvements in design and knowledge work. Historically, I've almost exclusively used Opus for slide decks, charts, and other visuals because I found it so much better at making design decisions and building graphics. If Sonnet can handle more of that work while consuming less usage, that would obviously be a major benefit.

It seems like Anthropic recognizes that a lot of people doing "regular" work, including accounting, use Sonnet for most of their tasks. I like seeing them put more attention into the documents, spreadsheets, and presentations we're actually working on every day.

I want to test it more, especially on the work I've usually reserved for Opus. But I think this is a huge update for anyone using Sonnet as their daily driver. Hopefully you'll notice an immediate difference like I did.

Numeric launched its Financial Data Platform on September 22, expanding beyond close automation into what it describes as an ERP replacement for companies with complex accounting.

The platform pulls financial activity from systems such as billing, payroll, and customer management software. It retains the original records, source documents, and supporting details rather than reducing each transaction to a journal entry. Numeric says that additional context gives accountants and AI agents better information for reporting and automation.

Accountants define rules for how transactions should be classified, allocated, and recorded, then review exceptions. Modules cover billing and revenue recognition, cash accounting, payables, fixed assets, leases, and month-end close. Reporting can use the underlying business details, such as revenue by customer, product, or region.

Companies can adopt the platform one module at a time while continuing to post entries back to NetSuite, then move toward a full replacement. Numeric is offering demos and has not published standard pricing for the platform.

Why it’s important for us:

I’d chalk this one up as extremely interesting, but TBD. I like the vision. We’ve just heard versions of it from plenty of vendors, and building an AI-native ERP is a lot different than announcing one.

For the same reasons I don’t think QBO is just going to disappear, I don’t think NetSuite is going to either. But I do expect major disruption over time. I have no reason to believe companies like Rillet, Numeric, and Campfire can’t be part of that. We’re already seeing it.

The focus isn't even really on the accounting. It's on the data that feeds into the GL. Historically, that data has been scattered across different systems, and getting it all into one place has been (and still is) a pain. By the time it reaches the GL, a lot of the useful detail is elsewhere.

So building from the ground up with better-connected data, like these companies are trying to do, makes a lot of sense to me. It’s just really difficult to do well.

I’ll definitely be watching this one. Cool announcement, compelling direction, but I’m not ready to overreact to it yet.

Trending News

OpenAI introduced a Decisions API powered by GPT-6 Luna for classifying content and routing work: This is the OpenAI response to Jev, which I covered last week. This could be a huge unlock for using AI far more cheaply. Good use cases could be anything with categorization or decision-making, such as matching transactions to a COA and routing work or documents.

Google announced Gemini 4 Argon, reporting stronger performance on finance, tax, and longer business workflows: Is Google back? Hard to know until the model becomes available for everyone, but at least it sounds promising.

KPMG formed a new group to build AI-native businesses through startup-style teams: So now the biggest advisory firms in the world want to be startups too? Ironic. Regardless, I actually like this idea. It's easier than ever to create custom solutions for businesses. Accounting firms are in an interesting position to be able to deliver on that, especially those who already offer advisory and who are leveraging AI.

Deloitte unveiled an AI-powered M&A platform it says is already used across more than 1,000 client engagements: AI in the M&A lifecycle makes a lot of sense to me. AI is good at a lot of the research, due diligence, and advisory that Deloitte's M&A team would handle.

SpaceXAI launched Team Bots in public beta, sharing files, instructions, and tools across teammates while keeping individual conversations private: They're actively trying to make Grok Bot more useful for businesses. I think they've seen some demand for that specifically, so I expect more in this area.

AICPA launched a Council on AI Risk in Tax with former IRS Commissioner Danny Werfel: Accounting regulators have really dropped the ball when it comes to AI. No useful guidance or opinions for the longest time, and still largely nothing to rely on. I'm also very skeptical that whoever is issuing guidance will be knowledgeable enough on AI considering how quickly everything is moving.

OpenAI launched ChatGPT Teams and shared Team Tasks for recurring work triggered by schedules or connected-app activity: A big theme of OpenAI's DevDay was working in the cloud and amongst teams. I like the idea of shared spaces for scheduled tasks and specific work. Downside is these use workspace credits instead of existing account usage.

OpenAI announced plans to integrate specialist dots with Microsoft Agent 365’s governance and security controls: Anyone building agents in Microsoft 365 should pay attention. It sounds like this will let teams manage OpenAI’s specialist agents through the Microsoft controls they already use. This announcement doesn't have a timeline yet.

Grok Bot added a Finance integration for connecting bank, card, and investment accounts: This is definitely centered more around personal finance than what firms handle. But firms' clients could absolutely be using something like this now or in the near future.

Put It to Work

Claude Docs, Slides, and Design are definitely a big game changer for working in documents and slide decks. The collaborative nature of working with your team and Claude in the same files is extremely enticing. I realize there are other document storage considerations for firms, but this could also be a great way of working on draft files and workpapers before converting final versions to Word Docs and PowerPoint.

Weekly Random

Would you notice if the person on your video call wasn’t a person?

Tavus introduced Griffin this week, an AI that generates a face, voice, and movements during a live video conversation. It can listen while talking, react to your expressions, and handle interruptions instead of waiting politely for its turn.

In Tavus’s study, 26 of 54 participants thought they’d spoken to a real person after a one-minute call. They’d been told they were meeting another participant, so they weren’t exactly hunting for an AI.

I watched the videos, and it’s incredibly easy for me to tell it’s AI. But I also spend an objectively unhealthy amount of time thinking about AI. Someone who doesn’t might have a very different reaction.

I can see good uses for this. Practicing a difficult conversation or getting help from something that can actually see what you’re doing sounds useful. Talking naturally instead of figuring out how to prompt a chatbot would be nice too.

But what’s the main point of making it look so much like a human? To make people think they’re talking to one?

Tavus says it’s working on features to disclose that it’s AI before a broader release. But building something convincing enough to fool people, then putting it in front of people who might not recognize it, still feels weird to me.

I’m not sure yet if I need my AI to have a face. Especially one that looks human. I already have Claude and ChatGPT ignoring my instructions. Giving it the ability to make eye contact while doing so feels unnecessarily disrespectful.