Harvey closed a $550 million funding round on September 9 at a $15.5 billion valuation, the legal-AI startup disclosed, capping a year in which its push into autonomous agents briefly sent gross margins careening from positive 50 percent into negative territory.
The swing happened fast. When Harvey launched Agent Builder in March, the low-code tool that let lawyers design multi-step autonomous workflows, token consumption exploded by roughly twentyfold through the first half of the year, according to Bloomberg reporting published in September. Gross margins, which hovered around 50 percent in January, had collapsed to negative 50 percent by June. The company scrambled through summer to build infrastructure that could handle the load without bleeding cash, eventually recovering to positive margins by late August.
Harvey declined to share specific financial details with Bloomberg, but the contours of the crisis were visible in company posts and executive interviews scattered across the spring and summer.
"Serving the most capable models as agents at scale is extraordinarily expensive," Gabe Pereyra, Harvey's president and co-founder, wrote in a June technical post explaining the startup's decision to build its own cloud agent runtime from scratch.
The Product That Broke the Model
Agent Builder, announced March 9, represented a conceptual leap beyond Harvey's earlier AI assistant tools. Customers could design long-horizon agents in natural language to tackle due diligence, contract drafting, and term extraction. Within days of launch, the platform was processing more than 400,000 agentic queries daily and running 25,000 custom workflows, the company said.
Harvey had also shipped deep Microsoft 365 integration just days earlier on March 4, including an "Agentic Word" add-in for redlining and structured edits that promised end-to-end legal workflows inside the Microsoft suite. On March 5, the startup upgraded to GPT-5.4. The confluence of new features arrived almost simultaneously, and usage spiked in ways the company hadn't modeled.
Winston Weinberg, Harvey's CEO, declared in a May company post that "the legal industry is now well past AI as an assistant and officially in the era of legal agents." The enthusiasm was warranted, perhaps, but the unit economics told a different story.
When Tokens Outrun Revenue
Harvey processed roughly 1 trillion tokens in January. By May, that figure had ballooned to a run rate of 12 to 13 trillion tokens per month, Weinberg disclosed on a June podcast. Bloomberg later cited a person familiar with the matter who pegged the year-to-date increase at twentyfold.
The problem was straightforward: token consumption raced ahead of revenue tied to Harvey's seat-based pricing. Lawyers were using the agents more aggressively than anticipated, and each agentic workflow burned through far more tokens than a simple query-response interaction. Harvey had raised $200 million at an $11 billion valuation on March 25, two weeks after Agent Builder launched, ostensibly to "scale agents across law firms and enterprises." But scaling agents, it turned out, meant scaling costs faster than income.
The margin collapse forced an engineering reckoning over the summer.
The Summer Rebuild
Harvey shipped three overlapping fixes between June and August, each designed to claw back profitability without sacrificing capability.
On June 1, Pereyra detailed a homegrown cloud agent infrastructure that routes tasks across multiple models from Anthropic, OpenAI, and Google, matching capability to cost. The company claimed it achieved "empirically 3 to 5 times cost reductions versus a frontier-only approach." Translation: not every legal task requires the most expensive model on the market.
In mid-August, Harvey launched Harvey II, a system that ties usage and cost to specific legal matters or projects. Administrators gained visibility into AI spend per engagement, a feature that gave clients more control over runaway token bills. Two days later, the startup announced Tenet, its first post-trained legal model built on Moonshot's Kimi K3 open-weight base in partnership with Fireworks. Harvey positioned Tenet as delivering "frontier-level" performance on legal tasks at open-source cost.
The Next Web reported in late August that the Kimi K3 choice was explicitly cost-driven, aimed at reducing per-token expenses while cutting vendor lock-in risk. By the time September arrived, margins had recovered to positive territory.
A Broader Industry Reckoning
Harvey's margin swing reflected a pattern rippling across vertical AI startups. Bloomberg Law reported in August that Thomson Reuters, whose CoCounsel product competes directly with Harvey, was shifting to an unnamed open-source base to reduce reliance on Anthropic and OpenAI. Other companies, including Abridge, Decagon, Ramp, and Rogo, were exploring similar moves, Bloomberg reported in September.
TechRadar Pro and ITPro documented what they called the end of "tokenmaxxing" in September, describing enterprise clampdowns on uncontrolled token spend as agents' autonomous loops made cost visibility harder to track. The agents worked well, maybe too well, but they created financial exposure that many companies hadn't anticipated.
Harvey's customer roster includes law firms such as Willkie Farr & Gallagher, Gowling WLG, Clayton Utz, and Hengeler Mueller, according to May company posts. The startup said it crossed $400 million in annual recurring revenue and signed more than 3,000 customers as of September, self-reported figures disclosed via the founder's LinkedIn and trade-press coverage rather than audited filings.
Where Things Stand
Katie Burke, Harvey's chief operating officer, told Legal IT Insider in September that the company uses seat-based pricing "but with the option to do consumption" and is "building in the ability for administrators to pre-select model per task choices" for cost governance. The messaging suggested Harvey learned its lesson: give customers the tools to control spend before usage spirals.
The September funding round, co-led by Diffusion and Lightspeed, came six months after GIC and Sequoia co-led the March round at an $11 billion valuation. The jump to $15.5 billion arrived only after the margin recovery, a signal that investors believed Harvey had solved the unit-economics puzzle that agentic products inevitably create.
Pereyra's June post had framed the stakes plainly enough. Multi-model routing and homegrown infrastructure weren't optional features for companies running agents at enterprise scale. They were survival requirements.
