Claude Opus 4.7 Shipped Yesterday. It Scored 50% on Month-End Close.
Anthropic's newest AI just posted the highest accounting benchmark score ever recorded. It still scored 50% on month-end close. The machine they told you would replace your firm is currently closing books at coin-flip accuracy. Here's what that means for the work you're still carrying.
Anthropic rolled out their smartest AI ever yesterday.
It scored 50% on month-end close.
The machine they told you would replace your firm is currently closing books at coin-flip accuracy.
Most of the takes on LinkedIn today are missing the actual story. Let me walk you through it.
The test nobody is running on these announcements
DualEntry makes accounting software for real firms. They put Claude Opus 4.7 through an actual benchmark of accountant work. Not marketing demos. Not cherry-picked prompts. The work you did on Friday.
The scorecard:
- Categorizing transactions: 92% right
- Posting journal entries: 92% right
- Running a month-end close: 50% right
- Writing a financial report: 62% right
Opus 4.7 is the best any AI model has ever scored on this benchmark. It still flipped a coin on close and botched one in three numbers on reporting.
If you've personally closed a set of books, you already know why.
Why close is different
Categorizing transactions is pattern matching. Office Depot goes to Office Supplies. Uber on a Tuesday in April goes to Travel. Machines will eat that work eventually. Honestly, most of us won't miss it.
Closing is something else entirely.
Closing is spending two hours tracking down a $4,200 variance and finding it was a wire that hit on the 27th and got coded to the wrong intercompany entity. You knew where to look because you've seen this pattern on this client for eight years.
Closing is knowing that when the bank rec is off by a round number, it's not bank fees. It's a missed deposit. Every single time. You can't train a machine on that. You learn it by being wrong for a decade.
Closing is calling a client at 9pm to confirm whether a $180K wire is a refund, a capital call, or a bonus the founder forgot to mention. And then making the call.
Fifty percent. The AI flipped a coin.
What this actually means for your firm
I talk to firm owners every week. Usually $1M to $5M in revenue. Usually moved upstream into advisory and CFO work. Every one of them has the same question now.
"Is AI going to take my firm?"
No.
The firm down the street might take your firm. The AI won't.
When the 92% work gets cheap, the 50% work gets more expensive. A lot more expensive.
If a competitor can categorize 100,000 transactions in an afternoon for forty bucks of compute, your margin on basic QBO bookkeeping just disappeared. Clients will demand the price cut. You won't be able to hold the line.
But if you've spent five years building real depth in cross-border tax, or ESOPs, or construction accounting, or PE fund admin?
You just got a raise.
The compliance work is getting commoditized. The judgment work is getting more valuable by the month. That gap is the one you should be watching.
What to actually do
Two moves are available right now, and most firm owners are making neither.
First: stop carrying the work AI is going to eat.
You know exactly which clients I mean. The thirty 1040s you can't bring yourself to fire. The twenty basic bookkeeping accounts from back when the firm was smaller. They're generating $80K of revenue inside a firm with $220/hour overhead. The math stopped working two years ago. You just haven't done anything about it.
Those clients are worth real money to the right buyer right now. A cloud-native firm with a lean back office can absorb that book tomorrow and run 70% margins on it. They want it. You don't.
Second: start actively sourcing the work AI can't touch.
If you're the specialist, you should be seeing every deal in your lane that other firms can't serve. Every week. Every state.
Until recently, the infrastructure for this didn't exist. You couldn't find a specialist when you needed one. The specialist couldn't reach the 200 firms quietly turning down work every month. And nobody was going to pay a broker 10% to move a $40K client block.
That's the problem FirmLever Network was built to solve.
What's happening on the platform
207 member firms. Deals closing every week. Three from the last month that illustrate the point.
- Philadelphia, $190K/yr outsourced CFO client. Listed by a firm that had outgrown it. Three firms raised their hand inside of three hours. Claimed same day. Fair price. No broker.
- North Carolina, 20 bookkeeping clients as a block, $30K/yr. 19 firms expressed interest. An AI-native firm won it. A legacy firm got liquidity. Both sides walked away happier than they were a week earlier.
- Florida, U.S.–Australia cross-border engagement. Australian unit trust, super fund, missing FBAR filings. Completely outside the listing firm's lane. Posted it as a service referral. Four international tax specialists picked it up in days.
AI will never do that Australia deal. The client still needed someone to do it. The generalist didn't want to lose the relationship. The specialist needed to find the deal. The platform made the match in under a week.
Members list client blocks they want to unload. Post what they're actively trying to acquire. Send and accept referrals. Build private referral relationships that compound every month. No broker skim. No six-month exclusive listings that go nowhere. No fee on the transaction.
The actual question
AI is not taking your firm.
The question is whether the work you're still holding is the work AI will eat, or the work that will be worth more every year from now.
If it's the first, you need to move those clients. Soon. The window where they're worth real money is already narrowing.
If it's the second, you need to be findable. Every week, firms on our platform are trying to refer work to specialists and can't, because those specialists aren't on the network.
The market is running at app.firmlever.com.
No broker. No commission. No fee on the transaction. Just firm owners moving the wrong clients off their books and the right clients onto them.
See you on the inside.
Marc
P.S. — The 19 firms that raised their hand on the North Carolina bookkeeping block all thought they were early. The 18 who lost it are waiting for the next one. If you want to see listings before the rest of the market catches on, get in now.
