Legora seeks $10B+ valuation, doubling in four months
The legal-AI money race just escalated: Swedish startup Legora is reportedly raising at a valuation above $10 billion, and a new Berkeley benchmark argues the industry is measuring machine learning itself all wrong.
Swedish legal-AI startup Legora is seeking fresh funding at a valuation above $10 billion — roughly double the $5.6 billion it commanded just four months ago, according to the Financial Times. The company, founded in Stockholm in 2023 and formerly known as Leya, builds a collaborative AI platform for lawyers and has been on one of the fastest valuation runs in enterprise software: $675 million in May 2025, $1.8 billion in October, a $550 million Series D led by Accel that put it at $5.5 billion in March, and Nvidia's NVentures adding $50 million at $5.6 billion in April. Legora says it has passed $100 million in annual recurring revenue and serves more than 1,000 customers. The company has not commented publicly on the reported round.
The target matters as much as the number. A $10 billion-plus valuation would pull Legora nearly level with Harvey, the category leader that raised at $11 billion in March, turning the legal-AI race into a two-horse contest at unprecedented heights. It is also the clearest sign yet that application-layer AI — not just the labs building foundation models — is where the money is going: legal is arguably the most monetized AI vertical, with tools that bill against six-figure lawyer time and a customer base that pays for defensible accuracy. The pace is the story, though. Doubling your valuation in four months, twice in a row, prices in a future where AI lawyering becomes the default workflow — and leaves incumbents in traditional legal software increasingly exposed.
A new benchmark out of UC Berkeley's Sky Computing Lab found that plain in-context learning — no memory management at all — beats elaborate context-management systems at continual learning. Continual Learning Bench 1.0, released by PhD student Parth Asawa with collaborators from Snorkel AI, the University of Washington and Wisconsin–Madison, runs systems through expert-validated task sequences across six domains — poker, database exploration, epidemiology cohorts, sales forecasting and more — and measures "gain": the difference between a system's score with memory maintained versus reset between instances, isolating true learning from base-model strength. The first results surprised even the team: vanilla in-context learning topped the reward and cost frontiers, while fancier notepad-style memory systems failed in telling ways — storing correct answers but refusing to apply them, or forgetting earlier feedback after a correction. We covered the commercial bet on this problem earlier — Trajectory raises $40M for AI that learns without forgetting.
The broader point is that today's leaderboards measure static capability, not learning — a model is scored at a single point in time, as if it forgets everything between tasks. As agents move into jobs that require adapting to new tools, schemas and feedback, that blind spot is becoming a product problem, not just an academic one. Benchmarks shape what the field optimizes for; if CL-Bench catches on, memory and adaptation become first-class metrics instead of afterthoughts.
What to watch: whether Legora's round closes near Harvey's $11 billion mark, and whether context-management vendors respond to CL-Bench's results.
Legal AI valuations are doubling every few months — durable momentum or froth? Tell us in the comments.
