All issues
Issue 09
The new Claude is built for analyst-length work.
Anthropic ships Claude Fable 5, built for hours-long work, with a top finance-benchmark score. Its safeguard stumble is a vendor-governance lesson for funds.
Hi folks,
The one thing
Anthropic's Fable 5 is the first frontier model aimed at hours-long work
- Shipped 9 June, the first generally available model in the tier above Claude Opus. Anthropic's claim: "the longer and more complex the task, the larger Fable 5's lead", with a top score on Hebbia's Finance Benchmark (Anthropic).
- It is included on paid Claude plans at no extra cost until 22 June.
- The launch stumbled: a safeguard buried in the 319-page system card quietly degraded answers for users suspected of building rival AI. Developers found it within hours; Anthropic apologised and made it visible within a day (Fortune).
- One for the compliance-minded: all Fable 5 traffic is retained for 30 days for safety monitoring. Not used for training, access logged, then deleted, but zero-retention arrangements do not apply to this model (Anthropic).
What this means in plain terms:
- "Longer task, larger lead" is analyst-shaped work: data-room reads, covenant checks, memo drafts. The jobs that take an afternoon, not a prompt.
- The window to 22 June is the cheapest test a fund will get. Hand it one real multi-hour task and compare.
- The vendor lesson: silent behaviour changes are the one thing a regulated workflow cannot absorb. Ask your AI vendors what changes without notice.
In the mix
- The AI giants are racing to the stock market (CNBC)
- OpenAI confirmed its confidential filing on 8 June, a week after Anthropic's. SpaceX, which absorbed xAI, started trading 12 June at roughly US$1.77 trillion, the largest listing in history (TradingKey).
- Why it matters: the prospectuses will put audited frontier-lab economics on record for the first time.
- Apollo and Blackstone anchored a US$35B AI compute platform (Blackstone)
- With Broadcom, more than 20 gigawatts of compute for Anthropic and OpenAI through 2028. This is the second giant AI financing from these names in a fortnight, after the US$36B chip deal we covered last week.
- Why it matters: private credit's largest managers are now funding AI infrastructure at platform scale, not deal by deal.
- Lenders are asking what AI does to their software loans (CNBC)
- Private credit has lent heavily to software companies because their recurring revenue is dependable. AI loosens that: some customers can now build the tool themselves instead of renewing.
- The read from specialists at Ares and Man Group: not a collapse, a sorting. The work is telling borrowers that adapt from those that get disrupted.
- Why it matters: if there are software loans in the book, AI risk now belongs in the underwriting.
From my week
- Three hours, unsupervised
- I talked Claude through what I wanted, then let it build, end to end, for three hours. It stopped when the credits did.
- The shift: I checked in occasionally instead of staring at it. That changes what you can hand over.
- Walk-throughs beat write-ups
- Our calendar this week was mostly hands-on walk-throughs, sitting with a team getting set up on Claude. They worked better than any document we could have sent.
- The leverage is in the basics: knowing what is possible, then having the right connectors, instructions and skills in place.
- Jake built the guide we use to get people started: instructions for Claude.
Written by
Levercon builds the AI operating system for credit funds: we connect your data, deploy agents across origination and monitoring, and run the repetitive work end to end.
Get Started
Book a strategy call.
See where AI fits in your fund.
Walk through Levercon's custom AI operating system to see how it can benefit your fund.