Harvey closed a $550 million round to fund building its own foundation models, while Legora's reported chase of a $10 billion valuation remains unconfirmed. LEGALFLY launched a contract platform that routes agreements from drafting through signature and post-close tracking, and Legora's own benchmark of OpenAI's new GPT-6 Astra found a 40% gain on one workflow against just 3% across its full suite. The same week Google pitched its legal AI tool, the D.C. Court of Appeals struck a Deutsche Bank brief over four citations Google's own AI had invented.
On September 9, Harvey confirmed a $550 million round at a $15.6 billion valuation, Bloomberg reported, roughly the figure first reported when talks began in early August. The round is earmarked for Harvey to build its own foundation models, a shift for a company whose product has always been a legal-specific layer wrapped around other labs' models.
Lightspeed Venture Partners and Diffusion, a new firm co-founded by longtime Harvey backer Kris Fredrickson, co-led the round; Sapphire Ventures, Whale Rock Capital Management and existing investors also participated, per Bloomberg. It's Harvey's second valuation step-up this year, up from $11 billion in March. Legora, meanwhile, is reported to be seeking a fresh round at a valuation above $10 billion, Artificial Lawyer wrote on September 9, up from the $5.6 billion the company reached earlier this year. Neither the amount nor the valuation for Legora's round has been confirmed on the record.
Harvey raising capital to build its own foundation models is a bet that the model is where the value sits. Precisely and other CLM vendors made the opposite bet last week: that once frontier models are commoditized, the value sits in the policy and verification layer wrapped around whichever model does the generating. Both bets can't be the differentiator for the same client. A better proprietary model doesn't change who decides which of a legal team's contracts gets routed to outside counsel and which gets handled in-house. That's still inexpensive work done by expensive resources until something owns the routing, not just the drafting.
LEGALFLY launched Contract Intelligence on September 4, extending the Belgian-founded platform into what it calls a system of record for the whole contract lifecycle, from the first draft through the obligations a signed agreement creates months later. The pitch isn't a better reviewer. It's a system that notices a renewal deadline or an unmet obligation on its own, and routes it to whoever needs to act, instead of waiting for a person to remember to check.
The platform reviews a contract against a legal team's own playbook, flags key terms and risk, and routes it for approval and e-signature. After signing, it tracks expiry dates and notice periods so a renewal or termination window doesn't lapse unnoticed, giving a legal team one continuously updated view of where risk sits across every active agreement, according to Legal IT Insider's report on the launch.
This is the layer Flank's own clients describe wanting when they explain the problem in their own words: not a faster reviewer, a system that knows what to do with a contract once the review is done. Routing an NDA to the right template, escalating an actual risk to a human, remembering an obligation three months after signature, that's mechanical work that currently falls on a lawyer only because no infrastructure does it by default. A CLM vendor building that layer is a second, independent bet that the bottleneck in contracting was never how fast the first draft gets reviewed. It's what happens to a contract afterward, and whether anything is actually watching.
OpenAI released GPT-6 Astra on September 3, and Legora ran it against a financial-statement tie-out across 41 documents, catching all four errors the firm had planted, including a £500,000 gap in a revenue note. Legora reports Astra scored nearly 40% higher than its previous model on that specific workflow. Across its full Benchmark for Agentic Reasoning suite of legal tasks, the average improvement was about 3%.
Harvey separately reports better handling of unsupported assumptions and gaps in documents after testing Astra, per Artificial Lawyer's September 7 coverage, without publishing comparable figures. Both readouts come from the vendors' own testing, published through OpenAI's and Legora's blogs; neither has been replicated by an independent benchmark.
A 40% jump on one workflow and a 3% average lift are both true at once, and only one of them describes what a model upgrade does for most of what a legal team actually asks AI to do. If a platform's pitch leads with its best single-workflow number, ask for the average across everything it runs, not the number that made the press release.
On September 3, the District of Columbia Court of Appeals, the district's highest local court, struck an appellate brief filed on behalf of a Deutsche Bank subsidiary after finding four case citations that counsel admitted were invented by Google's generative AI search tool. The panel's opinion registered its own surprise that a filing from a "competent law firm representing one of the largest financial institution[s] in the world" contained fabricated case law, and separately said its own authority to sanction the lawyers involved was unclear.
The case, Douglas v. Deutsche Bank National Trust Co., arose from a foreclosure appeal brought by a self-represented homeowner against the bank. Two days later, on September 5, legal researcher Damien Charlotin's public tracker of AI hallucination cases in litigation logged its 2,022nd entry since April 2023, more than 1,300 of them from US courts. Google's own models have been named in roughly 10 of those cases against more than 100 for ChatGPT, and fewer than 160 have led to formal discipline, mostly fines between $1,000 and $5,000. The same week, Google was promoting its own legal AI tool to law firms as a complement to their work, per reporting picked up by the Daily Caller on September 8.
"A competent law firm representing one of the largest financial institution[s] in the world" filed a brief containing AI hallucinations.
The panel's opinion, Douglas v. Deutsche Bank National Trust Co.| Metric | Value |
|---|---|
| Hallucination cases logged (through Sept 5) | 2,022 |
| Of those, in US courts | More than 1,300 |
| Cases citing Google's models | About 10 |
| Cases citing ChatGPT | More than 100 |
| Cases resulting in formal discipline | Fewer than 160 |
| Typical fine range | $1,000 to $5,000 |
Deutsche Bank's outside counsel is not a solo practitioner cutting corners. It is, in the court's own words, a competent firm representing one of the largest financial institutions in the world, and it still put four invented cases into a filing that a human was supposed to check. The volume problem isn't confined to consumer tools or under-resourced practices. It shows up at every tier of the market, because checking whether a citation exists is inexpensive work, and it keeps landing on expensive resources who are stretched too thin to do it reliably at the pace AI now enables. A supervised system doesn't remove that check. It makes sure the check happens before anything reaches a filing, regardless of which lawyer signed it.
Harvey is spending new capital to own its models. LEGALFLY built a product that routes contracts instead of just reading them. Legora's own benchmark shows a smarter model moves one workflow a lot and the average barely at all. And a fake citation from a free consumer tool still reached a filing for one of the world's largest banks. Every story argues, in its own way, about where the value in legal AI actually sits. Only one of them, LEGALFLY's, bet on the routing layer instead of the model.
Take this week's stories together: one vendor is betting on the model, one is betting on the routing layer, one benchmark shows the gap between a headline number and the average, and one filing shows what happens when nobody's watching for errors, whatever the resource level. None of that is routing infrastructure a legal team owns for itself.
Outsource legal work to supervised agents and the question stops being which vendor's model is fastest this quarter. It becomes a mechanism: agents that draft and route against a team's own templates, terms and escalation rules, with a human reviewing the output before anything leaves the building. Inexpensive work stops being done by expensive resources, whichever model happens to be underneath.