Harvey's Margins Collapse from 50% to -50% as Agent Token Use Soars 20x
Bloomberg reported on Sept 22 that legal AI unicorn Harvey's gross margin slid from 50% to -50% by June, as agent features drove token usage up 20x on rented OpenAI and Anthropic models, with inference costs devouring margins.
Confirmed
- Bloomberg: margins fell from 50% to -50% within six months, as of June.
- Agent features caused roughly 20x growth in token usage and sharply higher costs on rented OpenAI/Anthropic models.
- Harvey's response: building its own models on open-weight bases to cut costs and reduce dependence on OpenAI and Anthropic.
- Bloomberg also notes legal AI startups broadly adopting open models.
Unconfirmed
- @thedealdirector argued Harvey's early product was mediocre with low usage, enjoying high margins via seat-based pricing to lawyers buying 'because AI' — personal opinion, not Bloomberg reporting; the post also sparked discussion of Legora's margins.
Why it matters
- Harvey's margin collapse is a landmark case of runaway agent inference economics, warning application startups reliant on third-party model APIs.
- @jkubicki adds that law firms may likewise adopt open models to reduce reliance on vendors like Harvey and Legora, meaning cost pressure could propagate in both directions along the value chain.
2026-09-22 ~ 2026-09-22 · 5 related posts
Primary sources
- [source] Legal AI startups like Harvey turn to open models to cut reliance on Anthropic and OpenAI — jkubicki · 2026-09-22
- Harvey's gross margin plunged from 50% to -50% in 6 months, turns to open-weight models — nmasc_ · 2026-09-22
- Harvey's gross margin fell from 50% to -50% as agent token use spiked 20x — SumitGup · 2026-09-22
- [source] Harvey's gross margin sank to -50% as agent token usage surged 20x, Bloomberg reports — FinanceYF5 · 2026-09-22
- Harvey's gross margin allegedly swung from +50% to -50% as lawyers actually started using it — thedealdirector · 2026-09-22