Lock the firms, tune Beijing's open weights, sell to a publisher: legal AI vs OpenAI

A pension buys the week's most exposed winner

On 24 September 2026, Teachers' Venture Growth, the growth arm of Ontario Teachers' Pension Plan, put US$50 million into Harvey, stretching the 9 September round↗ co-led by Diffusion and Lightspeed from $550 million to $600 million at a $15.5 billion valuation ¹³⁴. That is up from $11 billion in March ⁴. The check sits at the bottom of TVG's $50-to-$250 million range, out of a $25.9 billion book spread across 45 companies ².

The supplier is now the rival

Harvey has passed US$400 million in annualized revenue, serves 20% of the Fortune 500 and is used by 80% of the 100 highest-grossing US firms ²⁴. Its engines are OpenAI and Anthropic models, and both makers now sell to lawyers directly ²¹¹. Astra for Law, launched 17 September, wraps GPT-6 in a legal index spanning more than 230 million URLs, with OpenAI engineers embedded at Sullivan & Cromwell, Ropes & Gray and Cooley building bespoke tools ⁵. Harvey's cost answer is Tenet, its post-trained model built on Kimi K3 from Beijing-based Moonshot AI ²³. TVG's Rick Prostko made that the thesis: Harvey "controls more of the intelligence behind its products. That is a durable advantage" ³.

Three ways out, all priced this week

Lock the firms in. Squire Patton Boggs chose Harvey for a global rollout on 22 September↗, with design-partner rights on coming launches ⁶; Eversheds Sutherland↗ has named Harvey its legal AI partner ⁷.

Re-rate. Legora is in talks, on one report, to raise $300 million-plus at roughly $8.5 billion pre-money↗, about 85 times its $100 million ARR; Harvey's $15.5 billion on $400 million-plus implies near 39 times ⁸. Three 2026 mega-deals absorbed 40.3% of the year's $3.35 billion of legal tech capital ¹¹.

Sell to a content owner. CMS and Dentons exited Noxtua in a Series C secondary worth more than $114 million↗ that made German publisher C.H. Beck the majority shareholder, both firms staying on as anchor clients ⁹¹⁰.

The other side

Winston Weinberg's defense: "the foundation models are competitors with every business on Earth," and law firms want multiple models, which is where the app layer earns its keep ². TVG's Olivia Steedman cites "the capabilities of the team, the product that they're building, the speed at which they're moving" ². The bear case is arithmetic. Astra for Law cleared 54% of one legal research validation set, not all of it, but every point OpenAI adds comes out of someone's multiple ⁵.

What to watch

Two readings decide the thesis: whether Tenet-class open weights hold Astra-level output at a tenth of the cost, as Weinberg argues they must ², and whether Legora's talks close at $8.5 billion or get marked down for platform risk ⁸. The fallback already has a price tag: $114 million, publisher in, law firms out.