Hello fellow keepers of numbers,

Several big announcements this week, including the new Opus 5. Anthropic released the new Opus as I was midway through writing this week’s newsletter, so I haven’t had much time to test it. But oddly, I did get to test its browser use capabilities in my demo video. Pretty impressive. Check it out.

Canopy ships another big set of AI features. They’ve copied some of the major tax intake vendors and have built an impressive set of features around their practice management system. KPMG partnered with OpenAI to build AI-native software for enterprises.

Plus, Claude Cowork will now record your screen and voice as you complete a task so it can create a skill to automate the work. I also demoed the new feature in this week’s newsletter.

The Latest

Anthropic released Claude Opus 5, a model it says delivers close to the performance of its top-tier Fable 5 at half the price. It is the company's fourth Claude 5 model release in under two months.

Anthropic positions Opus 5 as an everyday model for professional and knowledge work, and says it improves on the prior Opus 4.8 at research, document analysis, and numerical reasoning. Early enterprise users reported the biggest gains on tasks like financial research, data analysis, and due diligence review.

Opus 5 is available now on all paid Claude plans and through the API, priced at $5 per million input tokens and $25 per million output tokens, the same as the prior Opus 4.8. An optional Fast mode runs about 2.5 times faster at twice the base price.

Why it’s important for us:

Here’s my obligatory caveat for every new model release: benchmarks suck and I don’t trust them. I like to use the model myself and decide. That being said, this model came out as I was writing this, so I haven’t used it yet. The benchmarks compared to Opus 4.8 look pretty good. It’s a big jump in how good it is with knowledge work and business workflows.

Anthropic also noted that it’s significantly better at preventing prompt injections, which is a nice safety feature. Essentially, it can find out if a website or app is trying to do something malicious. This is important for those of us who use things like Claude Cowork or Claude Code to do research and/or pull data from a lot of sources.

I suspect it won’t be a noticeable difference in regular business workflows vs Opus 4.8. I think the most we can hope for with each new release is small improvements in intelligence and much better efficiency so we don’t destroy our wallets running up the usage.

Canopy launched Tax Workflow Automation, an AI-powered system that runs a tax engagement from document intake through preparation, signatures, payment, and delivery inside its practice management platform. The suite splits into three native modules.

Smart Intake automates document collection. It generates predictive document request lists, recognizes uploaded file types, and renames and organizes files, which Canopy says saves 20 to 50 minutes per client.

Smart Prep moves return preparation into the Canopy workflow. Selected documents pass to integrated partners where AI extracts data and populates returns, which Canopy says runs four times faster with up to 70 percent less manual work.

Smart Delivery closes the engagement out. It reads the finalized return PDF to auto-populate delivery details, then lets clients review, sign forms such as the 8879, and pay invoices in a single flow inside the client portal.

Canopy says every AI-generated output stays editable before it reaches the client, keeping a practitioner in control. Tax Workflow Automation is available now as a bundled Intake and Delivery package for Canopy customers.

Why it’s important for us:

Canopy is one of the most interesting vendors in the accounting space right now for two very opposite reasons.

Let’s start with the good because there’s a lot of it. The tax workflow automation looks awesome. They've essentially taken what the best standalone tax intake tools do and built it natively into the platform.

Create an organizer or upload a prior year one, build a questionnaire that dynamically changes the organizer based on how the client answers, and the organizer shows up for them immediately. The client drags and drops files, and the AI categorizes them. That's some of the coolest tax tech in the market right now, and Canopy is doing it inside their platform now.

Smart Prep is cool too because they've partnered with some of the leading AI tax prep vendors to power it. Smart Delivery can probably replace a few tools firms are paying for now.

Then there’s the other side. Canopy has thus far taken the approach of locking you into their software. Their API isn’t very robust. They don’t have an MCP for you to connect your data with Claude or ChatGPT. For me, that’s a major negative for any vendor right now.

If one of the features isn’t built as well as what a specialist vendor is doing, or something breaks, you’re relying on Canopy to fix it. You’re along for the ride with their roadmap and vision, whether or not it matches yours. And they’re always going to be trailing a little bit since they’re building versions of what the rest of the market is doing. They’re doing a great job at that right now, but it’s still a risk.

Jack of all trades, master of none, right? If Canopy is doing everything, you have to ask if any of it is individually the best in the market.

For some, betting on an all-in-one platform might be the right move. But we’ve also entered a world where software is moving at lightspeed. AI is changing things so fast, so I much prefer the tools that allow you to use your data however you’d like in whatever tool you’d like.

All that said, I really like what Canopy has shipped over the last few months.

KPMG announced a strategic alliance with OpenAI and was named an OpenAI Elite Partner, the top tier of OpenAI's partner network. The Big Four firm and OpenAI are building toward what KPMG calls a "headless" model of enterprise software, where systems like ERP and CRM keep running as the system of record but AI becomes the primary work surface.

In practice, employees stop logging into individual apps to get work done. Those systems keep handling data and transactions in the background, while AI agents coordinate the work across them through a single conversational surface.

KPMG says the alliance grew out of a build it did for OpenAI itself: an AI-native platform that unified OpenAI's own supply chain, fulfillment, and finance operations behind one conversational interface. After that project, the two decided to take the approach to market together.

Why it’s important for us:

This announcement is essentially KPMG and OpenAI agreeing to work on AI-native software for enterprises. “Headless” is a nerdy way of saying the software you log into stops being the point. The underlying data source matters less because the software on top is an agentic AI interface (like Claude Cowork).

Why open 3 apps and 7 Chrome tabs when I can open Cowork and have it pull all that same data into one place and synthesize it for me? The connectors I add in Claude matter. But where the data physically sits and how ugly it is becomes less important.

Building enterprise software from scratch with that assumption lands you at a very different place from current enterprise software. And your software can literally learn from itself and how you use the data.

There’s not much in this announcement beyond an idea. But I personally find it interesting, so I’m curious to see where this goes.

Anthropic launched Record a Skill, a feature in the Claude desktop app that lets users screen-record themselves doing a task, narrate the steps as they go, and have Claude turn the recording into a reusable skill it can run again. It sits under "Record a skill" in the + menu of the app's Cowork interface and replaces the old approach of hand-authoring skill instructions in a markdown file.

The launch follows a similar OpenAI feature, Record and Replay, which shipped for its Codex tool in June for Mac users. Record a Skill is available on Anthropic's Pro, Max, and Team plans and is not offered on the free tier.

Why it’s important for us:

OpenAI released a feature very similar to this about a month ago, and I honestly haven’t heard a thing about it since. Do people want this? I think it’s pretty awesome, but I’m also an AI nerd.

The best way to create a skill right now, in my opinion, is to do the task in Cowork, review it, and iterate with Claude until it’s done. Then, tell Claude to take that task and turn it into a skill based on everything you did together.

This is pretty similar. I actually think this is most useful right now for computer use skills. Remember the days when you’d create an Excel macro by recording yourself doing the task in Excel? It made note of where you clicked and then created the VBA script. Maybe people still do this. Cowork’s “record a skill” is similar.

Trending News

Filed launched OpenTax, an MCP server that lets Claude prepare a return and review a completed 1040: This is a tax engine. Seemingly free, which is cool. It's not going to file returns, but this could be really useful for extensions or estimates. Use Cowork to connect to the source docs.

CLA partnered with Digits to run its client base on Digits' AI-powered accounting ledger: This adds a lot of legitimacy to Digits for potential buyers. It's not only a very viable option now, but arguably more interesting than the native ledgers like QBO and Xero. It's still a bit of a risk since it's new and developing, but the upside is huge.

OpenAI gave ChatGPT Voice desktop control, letting users run and direct multiple agents in ChatGPT Work or Codex by voice: Voice has been a hot topic for a few years now with options like Wispr Flow, Superwhisper, and others. The alternative is what if your AI tool itself used two-way voice to listen and talk back? For those of us who are Iron Man fans, it's interesting to think about how we can have our own Jarvis.

Google made Gemini deck generation available in Slides, building an editable, on-brand presentation from a prompt plus your own Docs, Sheets, and PDFs: Somehow this is four months later than Claude in PowerPoint, despite Google trying to build their own AI to be compatible with their own tools. This does look interesting though if you're in the Google Workspace. Claude isn't quite as good at Slides as it is PowerPoint.

Cursor launched Cursor Router, which auto-picks the underlying model per request and, Cursor says, matched frontier quality at 60% lower cost: I love so many of the things Cursor are doing lately. Choosing the right model and effort for the right task is often one of the most annoying things about AI. Cursor is trying to entirely solve that problem by choosing for you. Early feedback seems good.

Google expanded Gemini Spark, its 24/7 background agent, to Google AI Pro subscribers in the US: My hopes aren't very high, but they're finally rolling out the Claude Cowork competitor. It's a tiny bit interesting for anyone on the Google Workspace stack, at least to test.

Google released Gemini 3.6 Flash, a cheaper, faster model that uses about 17% fewer output tokens than 3.5 Flash: Gemini has fallen off a cliff recently. Claude and ChatGPT are running laps around it at the moment. Instead of releasing their new Gemini 3.5 Pro model that's rumored to be strong, they give us a small update to the small, fast model. I'm worried Gemini is becoming irrelevant, if it wasn't already.

GAO warned the IRS is being flooded with AI-generated public comments on proposed tax rules, with no policy for handling them: What an all-time "no shit" moment. Have they checked the comments on Twitter or LinkedIn lately? My LinkedIn inbox is a cemetery of terrible AI-written messages. Maybe if our governing bodies would wake up, we'd have some useful guidance around AI by now...

Xi Jinping opened Shanghai's World AI Conference and pitched a Shanghai-based global body to write the world's AI rules: Whether they're competing with the U.S. to win the AI race or not, it's another noteworthy figure calling out for global AI rules. Hopefully this is a step in the right direction toward worldwide collaboration (yes, that felt stupid to type).

Put It to Work

Now that Claude can watch me do things on my computer, I think it’s finally realizing just how dumb I actually am. Nonetheless, this week I used the new Cowork “Record a Skill” feature.

My tokens are very valuable. I can’t be wasting them on ridiculous things. Which is exactly why I made Claude create a skill to get the weather for the next day for Houston, Phoenix, and Amsterdam, and log it in a Google Sheet. Very valuable stuff.

I also learned along the way that computer use isn’t available on the Claude Team plan right now. I suspect that’s a safety feature. But it meant my initial plan for the skill was adapted mid-video. It still worked. The new feature in Cowork is impressive if you’re the kind of person who wants to let AI watch you do work.

Weekly Random

OpenAI disclosed this week that two of its own models hacked Hugging Face, the site where the entire AI industry stores and shares its models.

OpenAI was testing how good the models are at hacking. That test runs on a walled-off computer with no path out to the internet, and it's scored, which means there's an answer key sitting somewhere.

GPT-5.6 Sol and an unreleased model figured out it was being tested. They decided the best way to get the answer correct was to find the answer key. The models figured out how to break out of their sandboxed environment and onto the real internet. They then found a flaw in Hugging Face's software that nobody knew about, so no fix existed. The models paired it with a stolen username and password. That got them inside the live systems Hugging Face actually runs on.

Hugging Face's security team caught it and shut it down on July 16. OpenAI didn't connect the break-in to its own testing until five days later, so the victim figured it out before the company running the experiment did.

I think it’s safe to assume at this point that if a top-tier AI model like Fable 5 or GPT-5.6 wants to hack software, it’ll be able to do so. That’s honestly pretty terrifying.

This is an example of a hack that was somewhat harmless. But it easily could’ve been something more malicious. We’re putting a lot of trust in a small subset of companies whose bias clearly makes them flawed. It’s why so many are calling for better regulation.

Glad I could end us on a fun note this week.