Hello fellow keepers of numbers,

It’s good that I don’t work in public accounting anymore because I’m a nerd and get way too excited when new things drop. I would’ve missed all my 4/15 return deadlines. But for those who still fight the fight, congrats on making it out of it yet again.

Anthropic launched their new model, Opus 4.7. It’s not their most powerful model (Mythos is), and they’ve specifically tuned down the model so that it’s not good at cyber. Which is kind of concerning.

We also have several fun accounting-specific updates. Digits launched a new outcome-based pricing model for accounting firms. EY rolled out agentic AI within their EY Canvas audit software, and they’ve said they expect it will support end-to-end audit activities by 2028. And Modus raised a bunch of money to invest in accounting firms and roll out their proprietary agentic audit software.

Plus, a demo of the new Claude Routines.

THE LATEST

Anthropic launched Claude Opus 4.7. The company says the new flagship model outperforms Opus 4.6 across software engineering, multimodal reasoning, tool use, and agentic computer use, with the biggest gains on the hardest tasks.

Opus 4.7 introduces a new "xhigh" effort level that sits between high and max, giving users finer control over the tradeoff between reasoning depth and latency on difficult problems. Vision also gets a boost: the model can process images up to 2,576 pixels on the long edge, roughly 3x the previous resolution ceiling.

On cybersecurity, Anthropic says it deliberately reduced Opus 4.7's cyber capabilities during training compared to Claude Mythos Preview, which remains restricted to the Project Glasswing coalition. Opus 4.7 includes safeguards that automatically detect and block requests flagged as prohibited or high-risk cyber use. Security professionals who need access for legitimate work like vulnerability research or penetration testing can apply through a new Cyber Verification Program.

Pricing for Opus 4.7 stays at $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.6. Opus 4.7 is available immediately across all Claude products and the Anthropic API, as well as Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry. The API model ID is claude-opus-4-7.

Why it’s important for us:

I think this is the first time an AI company has publicly admitted to dialing back the power of their model on purpose. Opus 4.7 is an upgrade on 4.6 across almost every benchmark. But Anthropic said they’ve deliberately reduced the model’s cyber capabilities.

This is particularly interesting because some underlying nuggets highlight the fact that this is a brand-new base model. Meaning, it’s not an adjustment to 4.6, but rather the model was built from the ground up. So maybe this is the same family as Claude Mythos, but just tuned down so it’s not as powerful? Purely speculation on my part.

A few things that are worth knowing if you’re using this model day-to-day:

(1) It uses more tokens per response than Opus 4.6 did. Anthropic bumped up the usage limits for all plans to compensate. Hopefully the new ceilings are approximately 1-to-1 with what we had before.

(2) A few people from Anthropic have suggested defaulting to the new “xhigh” effort level. This is probably a fine default for us as well. However, I suspect a lot of tasks for accountants are much less cumbersome, so you can probably get away with “high” or “medium” effort level for things like reviewing email inboxes, writing, or formatting. Choosing lower effort levels will spare your usage limits, but might yield lower-quality responses in some cases.

(3) Instruction following got stricter. People within Anthropic are saying Opus 4.7 takes your prompt more literally than prior models. You should probably try to be a little more precise in how you’re writing prompts. According to the same Anthropic sources, this has resulted in the model being much stronger at business work like writing, creating slide decks, preparing financial models, etc. This is a huge win for us.

Digits switched its accounting firm pricing to an outcome-based model. Under the new structure, firms only pay Digits for clients where the AI platform achieves a "Zero-Touch Transactions" rate of 95 percent or higher. The metric counts transactions that are booked, reconciled, and reviewed before period close without any human involvement from the accounting team.

If Digits falls below that threshold on a given client, the firm pays nothing for that client. The company pointed to one early partner firm with over 100 clients where the team touched only 2 percent of transactions in December, giving Digits a 98 percent zero-touch rate across the book.

The outcome-based pricing is live now for Digits' accounting firm customers.

Why it’s important for us:

I love this move. They’re putting their money where their mouth is, and it shows the confidence they have in what they’re building.

I’ve heard a couple of people who said, early on, Digits was a little rough around the edges and could occasionally get categorizations wrong. That’s obviously to be expected for a new tool using state-of-the-art technology. For the firms worried about that, Digits has now removed a decent chunk of the risk. Of course, there’s still the risk of relying on a platform that’s creating some rework, even if it’s free.

My guess is Digits was finding it difficult to convert accounting firms from QBO, Xero, Sage, and others. It’s definitely a big ask for a firm to change their GL software. Intuit knows this because they still can’t get half the accountants on the planet to move off QuickBooks Desktop…

It feels like this is Digits’ response to those concerns. It’s a pretty good one.

There’s clearly more to bookkeeping/CAS than just categorizing transactions. But this is still a pain point for a lot of firms. Digits needs to get everything else right as well if they want to grow in the space and remain a legitimate competitor to QBO, Xero, Sage, and others.

EY launched enterprise-scale agentic AI across its global audit practice this week, with the technology embedded directly into EY Canvas, the firm's single global Assurance platform. The rollout reaches 130,000 Assurance professionals, 160,000 audit engagements, and 150 countries. EY Canvas already processes roughly 1.4 trillion lines of journal entry data every year.

The agentic AI is built on Microsoft Azure, Microsoft Foundry, and Microsoft Fabric, with a multi-agent framework that orchestrates audit tasks, dynamically addresses identified risks, and pulls from continuously updated auditing and accounting guidance. EY says the goal is to embed AI across every phase of an audit, improve risk assessments, and reduce administrative burden on clients.

EY emphasized that the approach maintains "the fundamental role of human judgment, skepticism, and insight" throughout the audit process. The firm expects the system to support all end-to-end audit activities by 2028.

To support the rollout, EY has built a global training program to upskill every audit and technology risk professional in 2026, with both immersive and in-person components.

Why it’s important for us:

There’s a line buried in here that should probably be the headline. EY expects their agentic AI to support all end-to-end audit activities by 2028. We should address this for a second.

This is less aggressive than PwC’s timeline, which was within calendar year 2026. But this is now two releases in the last 6 months by Big 4 firms saying they plan to automate audit activities. Are these marketing ploys disguised as tech announcements? Possibly.

We should probably take this seriously though. It’s not implausible that a good chunk of audit activities can be automated by AI. A lot of it is collecting data, ticking and tying, and review. The professional judgment is still where humans are important.

Beyond that line, the announcement is a great signal for the rest of the industry. You don’t need EY’s budget to do this. The small firm version might be Claude Cowork, Claude Code, OpenAI’s Codex, or a handful of targeted agents wired into your existing workflows.

What you’re seeing in this announcement is EY capitalizing on its enormous amount of data. To take a mini victory lap for a second: I (and many others) have been saying for 2-3 years now that firms need to leverage their data. Figure out where it all sits, aggregate it into sources that are easily accessible by AI, and keep it clean. And where you’re not capturing data (e.g., lots of firms still don’t record their meetings), start doing it ASAP.

EY’s initiative is possible because they’re connecting AI agents to their audit data, guidelines, and robust examples. Firms that have been doing the work to aggregate and clean their data are in a great position to capitalize. Firms that haven’t need to start now.

Modus raised $85 million in combined Seed and Series A funding led by Lightspeed Venture Partners. The company describes itself as an AI-native audit technology platform and holding company that invests in and partners with audit-first accounting firms, then equips them with its AI platform rather than building a standalone firm from scratch.

The stated goal is to preserve the acquired firm's heritage, people, and client relationships while improving audit quality, efficiency, and scalability. Lightspeed has called Modus the first AI-native audit technology platform.

Since launching in June 2025, Modus has invested in a Top 200 accounting firm with more than $30 million in annual revenue. The company says the firm is on track to more than double its organic growth rate in 2026 after deploying the Modus platform.

The new funding will be used to accelerate technology development and support additional firm investments.

Why it’s important for us:

This is a hybrid model that we’re seeing pop up in several industries, including public accounting. I think the strategy is really interesting for both parties.

For Modus, it differentiates them from other AI-native software companies that run the risk of being eaten by Anthropic, OpenAI, and Google. Their long-term value comes from the success of their investments in firms.

For firms, they get access to technology and talent they likely wouldn’t have without the investment. Most firms aren’t going to be able to build and implement AI agents on their own. It’s not like it’s something you can just learn quickly and implement in between audit seasons. It requires a lot of strategy, effort, and change management. For a firm owner who sees the AI wave and doesn’t want to become a tech CEO at 58, a Modus-style deal probably makes a lot of sense.

That said, I don’t want to pretend this skips the pain. Every headache a firm owner would hit rolling out AI themselves is still waiting on the other side of a Modus deal. Training, workflow redesign, new software implementations, partner buy-in, staff buy-in, client communication. It’s all still a problem.

I don’t know anything about the Modus audit software. So I’m going to remain skeptical on ROI until I hear more. But this is an interesting alternative to PE. It sort of “rhymes” with PE, but it’s different. Right? I don’t know.

I’m generally optimistic about the large influx of money and tech into the accounting industry. In the long run, it’ll likely lead to much more innovation than we’ve seen over the last few decades.

TRENDING NEWS

Anthropic launched routines in Claude Code, a research preview that runs saved prompts and repos on a schedule, via API calls, or on GitHub events: Basically the same thing as scheduled tasks, but they can also be triggered by actions from other software via an API. This can be used with any tool that has a native connector in Claude or an MCP.

Anthropic redesigned the Claude Code desktop app around parallel agent workflows, with drag-and-drop panes, a session sidebar, and an integrated terminal: Quality-of-life upgrade if you live in Claude Code day-to-day. The sidebar alone is worth it if you've been juggling five sessions at once. You can have several windows open at once now as well.

OpenAI released the Codex app for macOS with native computer use, bringing agentic coding and screen control to the Mac: Codex is the Claude Code equivalent for OpenAI. Computer use still sucks right now. It’s slow and not nearly as smart at completing tasks when you compare to other use cases. But the Codex functionality for using the desktop in the background while you also work is great. Many people are really sleeping on Codex right now, like they were previously with Claude Cowork / Claude Code.

Perplexity pitched Perplexity Computer as your personal CFO after adding a Plaid integration for full view of your personal finances: "Personal CFO" is doing a lot of work in that pitch. Regardless, this is a look into what the future of personal finance could be.

Juno raised a $12M seed round to scale its AI tax prep automation platform for small and mid-sized firms: They’re definitely one of the main players in a market that’s getting crowded quickly. Interesting because the firm from which this software was born has been using it internally, so it’s battle-tested.

PwC plans a major overhaul of its global consulting business, restructuring the practice after a turbulent stretch for Big Four consulting: Consistent with the "no opting out of AI" push from PwC leadership. Not sure how the partners are going to react to this, but it’s the same conversation within all of the top public accounting firms right now.

OpenAI unveiled GPT-5.4-Cyber, a cyber-permissive variant of GPT-5.4 available only through its Trusted Access for Cyber program for verified security pros: This is their response to Anthropic’s Claude Mythos and Project Glasswing announcement. Not nearly as cool a name.

Perplexity launched Perplexity Computer as a native Mac app, giving users direct desktop access to the agent: Performance and cost of Perplexity Computer just aren’t there yet based on my testing. This is Perplexity still trying to capitalize on the OpenClaw movement. It could work, but they’re fighting an uphill battle.

Anthropic launched Claude Design, a research preview that turns a text description into prototypes, slides, and one-pagers, with export to PDF, PPTX, URLs, Canva, and more: This is an awesome announcement, but TBD on results until I can test it. Could be game-changing for client decks, one-pagers, internal memos, and branded deliverables.

PUT IT TO WORK

My entire life is a routine. I wake up tired, I do a lot of work, my children cause chaos, I annoy my wife for a few hours, and then I go to sleep. So when Claude launched routines, I got pretty excited. 

Turns out other people have more productive routines. And you can automate them with Claude Code now. Below is a demo of a routine for a simple client onboarding workflow (plus a bonus skill).

Claude Routines for Onboarding Demo - Watch Video

WEEKLY RANDOM

So I heard of this really great investment opportunity. We can all get in early. Hear me out. What if you could buy GPUs to train AI models from the same people who ran a shoe company into the ground?

I know what you’re thinking. Your first thought is “what’s the maximum amount of money I can invest?” But your next thought is “how does a shoe company have anything to do with AI?” Well, look here. First of all, “GPUs” rhymes with “shoes.” Obvious. Next, people who work in data centers wear shoes. Except for the robots…huh, what’s that? Robots are going to autonomously run data centers? Shit.

Check out Allbirds’ new mission, NewBird AI. Surely this has nothing to do with their stock price subsequently skyrocketing 600%. Definitely not.