Harvey AI's gross margins swung from negative 50% to positive in a single quarter as agentic token use spiked 20x

Harvey swung from negative 50% gross margins in June to positive margins within a single quarter, despite token consumption doubling month-over-month, according to founder Gabe Pereyra.

The company hit $400M ARR—roughly $33M monthly—meaning those June margins cost Harvey approximately $16M in that month alone. The margin collapse came as agentic token use spiked 20x over the year. Lawyers using Harvey's legal AI shifted from simple single-shot model calls to reasoning-powered agents that fan out across entire firm data: all policies, records, related matters. That architectural change generated vastly more tokens at a time when the company was still on seat-based pricing.

Harvey's recovery was mechanical. The company shifted token spend toward cheaper models, fine-tuned an open-weight model (called Harvey Tenet), and built internal infrastructure for customers to monitor and cap spend. Pereyra framed the negative margin period as a deliberate customer-first choice: pushing customers onto consumption pricing before they were ready would have "hurt our margins" less but served them worse. Instead, Harvey ate the cost while transitioning customers gradually.

Sequoia led Harvey's Series C at a $15.5B valuation in September, announced after the June margin crisis. The timing suggests Sequoia either had visibility into the negative margins or was comfortable with the burn profile—Harvey has raised $500M and a single month of negative 50% margins at $33M monthly revenue is manageable in that context.

The legal software industry runs at much higher margins: DISCO, a public e-discovery provider, operates at 75% GAAP gross margins. Thomson Reuters' legal professional segment reports nearly 50% adjusted EBITDA margins. Even accounting for the cost of token-based reasoning, Harvey needs to climb significantly to reach software-business norms.

The tailwind is brutal pricing on frontier models. Grok 4.7 launched yesterday and proved particularly cheap for legal tasks, scoring well on Harvey's own legal agent benchmark. Anthropic's Opus 5 rolled out today. With model costs dropping and quality climbing, cheaper routing and post-training become viable margin levers.

The debate now is philosophical. Max Junestrand of Legora argues that training proprietary models makes no sense when frontier models drop in cost and quality faster than custom work can keep pace. Harvey is betting that a mix—frontier models for some tasks, fine-tuned open-weight models for others—will be the right balance. Both positions assume token costs continue falling, which recent model launches support. The real test is whether Harvey can hold customers through pricing model transitions and reach 50%+ gross margins within the next year.