Issue 01
Claude's finance benchmark just jumped.
Opus 4.7 shipped with real finance numbers. BlackRock signals Aladdin for private markets. OpenAI acqui-hires Hiro. Oracle and BlackLine ship agentic AI for finance on the same day.
Hi folks,
I spoke to two Australian fund managers this week. One of them said: "We still haven't exactly landed on what we need to do. We just know we need to do some things." The other said: "We should probably all just stop and spend the day thinking about it. Could get that day back in that time, but we're so busy we haven't done it." If either of those is how the AI question feels from your seat, welcome. This brief is for you.
The one thing
Claude Opus 4.7 shipped this week (16 April) and the finance numbers are real.
- The model scored 64.4% on the Vals Finance Agent benchmark -- a test of real financial-analyst tasks like reading filings, building models, and extracting metrics from structured documents. Six months ago the best model in the world scored 55.3% on the same test. (Vellum benchmarks)
- Vision got a serious upgrade too. Opus 4.7 now reads full-page PDFs and scans at 3.75 megapixels (up from 1.15). That is the jump from "kind of works on clean digital text" to "actually reads the loan agreement you scanned last week."
- It is better at long-running agent tasks with memory -- think a covenant monitor that holds state across a quarter and flags breaches as new borrower reports land.
What this means in plain terms. None of this lets you hand off a live trade. But for an analyst, the following are now tasks where Opus produces a usable draft more often than not:
- First-pass IM read on a new deal, with the key financial snapshot pulled out
- Covenant and MAC-clause extraction from a credit agreement, with the clause and page cited
- A quarterly covenant-compliance summary from a borrower's PDF reporting pack
- A first draft of an LP letter or commentary
The value shift is not "AI does the task". It is "analyst time per task drops 60 to 80% with a review step at the end." For any fund running meeting-prep, modelling, or monitoring workflows, it is worth spending an afternoon testing against real documents from your own book before deciding what to build or buy.
In the mix
- BlackRock's Aladdin is coming for private markets. (BlackRock Q1 release, 14 April)
- Monday's Q1 earnings made "Aladdin for Private Markets" an explicit 2026 priority.
- The combined Preqin (BlackRock acquired 2025) and eFront (acquired 2019) integration is now live.
- Quick clarifier, because I got asked: Aladdin itself is not AI in the ChatGPT sense. It is portfolio-management and risk software built on statistical models, Monte Carlo simulations, and BlackRock's aggregated market data. It has run the world's biggest bond portfolios for 30 years.
- BlackRock has been bolting AI "Copilot" tools onto Aladdin over the last year.
- Why it matters in an AI brief: this is how AI arrives at scale in funds. Not as a standalone AI tool but baked into the incumbent platform you already pay for (or soon might).
- Why for AU GPs: you are almost certainly too small to buy Aladdin, but your LPs will start expecting Aladdin-grade reporting from you. That is the ratchet to watch.
- OpenAI acqui-hired Hiro Finance (14 April). (TechCrunch)
- Hiro was an AI-powered personal-finance planning startup.
- Team and IP move into OpenAI's enterprise division.
- Hiro sunsets its product on 20 April, deletes data 13 May.
- Why it matters: frontier labs are now buying finance verticals, not partnering with them. Expect a finance-shaped product inside ChatGPT Enterprise in quarters, not years.
- Oracle and BlackLine both shipped agentic AI for finance on the same day (14 April). (Oracle, BlackLine)
- Oracle added agents for treasury, trade finance, credit and lending inside their corporate-banking suite.
- BlackLine launched "Agentic Financial Operations" -- a governance layer for LLM outputs in finance and accounting.
- Why it matters: incumbent enterprise-finance vendors are racing each other. AI is coming into your stack whether or not you chose it.
From my week
- Claude can now run agents in the cloud on a schedule. Anthropic calls it Routines.
- One of the fund managers I spoke to this week told me he wanted "technology that could automatically pull down all those announcements and at the end of the day send you a PDF, look at it like the daily newspaper."
- Routines is the first plausible version of exactly that.
- It runs on Anthropic's cloud, on a schedule or when something specific happens, against your existing Claude subscription.
- Your laptop can be off.
- The setup is still a bit technical but the shape of the thing is approachable.
- Enterprise permissions are still the tax.
- I got Routines working on my personal Claude account in twenty minutes.
- Same thing on my enterprise account took three hours of wrestling admin settings.
- If your fund is on an enterprise plan, brace the IT team before you start.
- Discovery signal of the week.
- Every conversation I have had with a fund this month surfaces the same line: "I know I need to use AI, I don't have the time to work out how."
- Five different people now.
- Worth saying out loud because if you are feeling that, you are not alone. It is not a you problem.
Stat of the week
Vals Finance Agent benchmark
64.4%vs55.3%
Claude Opus 4.7 vs the best model six months ago. Worth watching this number over time. When it hits the mid-80s is when "AI did the task" stops being an overstatement. (Vellum)
Cheers,
— Cara
Levercon (previously Fund OS) is what we are calling this. I am talking to fund managers and their analysts about how they actually use AI in their work. Forward this if useful. Reply if there is something you want me to dig into next week.