Harvey

Harvey is a US legal-AI startup, founded in 2022, that builds large-language-model tooling for law firms and corporate legal teams, including document drafting, legal research, case analysis, and due diligence. A September 2025 Fortune profile, headlined around a $100 million figure, described the company's stated aim as wanting to "kill the billable hour." Later reporting places Harvey's annual recurring revenue near $100 million during 2025 and its private valuation at $3 billion as early as February 2025 (Source: techstartups.com). The company was founded by chief executive Winston Weinberg, previously a litigator at O'Melveny & Myers, and co-founder Gabe Pereyra, previously a researcher at Google DeepMind and Meta, out of experiments with early versions of GPT-3 (Source: techstartups.com).

Snapshot

Valuation

Revenue

Total funding raised stands above $1.2 billion, from investors including Sequoia Capital, Andreessen Horowitz, GIC, Coatue, Kleiner Perkins, Conviction and Elad Gil; a round on the terms discussed in August 2026 would take the total past $1.7 billion (Source: techstartups.com).

Overview

The Harvey platform is an LLM-based system for legal workflows, covering legal research, document review and drafting, contract analysis, due diligence and compliance work, and has moved into multi-step workflows through custom AI agents (Source: techstartups.com).

Two self-reported scale figures are on the record and do not match, and both are the company's own. In August 2026 Harvey said more than 100,000 lawyers across 1,300 organizations use its software and that more than 25,000 custom AI agents had been created inside customer organizations (Source: techstartups.com). The company website states 200,000+ professionals, 2,400+ law firms and in-house legal teams, 75+ AmLaw 100 firms, 70+ countries, and an average of 25+ hours saved per month (Source: harvey.ai). Its disclosed customer roster includes Latham & Watkins, A&O Shearman, O'Melveny, Comcast, Verizon, Deutsche Telekom, Reed Smith, Syngenta, Repsol, Cuatrecasas, Adecco Group, CMS, and Burges Salmon.

Harvey relies on frontier models from Anthropic and OpenAI as its underlying capability substrate, with Claude Opus 4.5, Claude Opus 4.6, and GPT-5.3 Codex cited among the models in use.

Spectre internal agent system

Spectre (named after a Dota 2 character) is Harvey's internal agent system. It autonomously handles engineering and, increasingly, non-engineering work, and is triggered by system monitoring of incidents, bug reports, customer feedback, and Slack messages. In his April 2026 essay, Pereyra described Spectre as "the beginning of a company world model: a live picture of what is happening inside Harvey and what needs to happen next," noting that most of what it does is not triggered by a human prompt; it monitors and triggers work itself.

Business model and market position

The Fortune profile (September 3, 2025) characterized Harvey's strategic positioning as wanting to "kill the billable hour": automation of associate-level tasks reduces hours billed, which Harvey frames as pushing law firms toward value-based pricing. Pereyra's April 2026 essay situates legal automation alongside AI's effects on call centers, junior engineering, and junior consulting.

Harvey competes with several categories of legal-AI offering. Thomson Reuters CoCounsel is the incumbent offering, also Claude-powered (see the Anthropic / Thomson Reuters case study), against which Harvey positions itself as the upstart. Springbok AI, Cleary Gottlieb's in-house platform, represents the alternative in which law firms build their own systems. Harvey also competes against general-purpose Claude and GPT, against which it differentiates on specialization, workflow integration, and compliance and confidentiality guardrails. Legora, a Swedish startup, is a further direct competitor, and the legal-technology incumbents Clio and Relativity are building their own AI capabilities (Source: techstartups.com). A Law360 Pulse article (October 22, 2025) reported that clients are driving AI adoption in law firms.

Pereyra "Legal Is Next" essay (April 2026)

In "How Autonomous Agents Will Transform Legal" (April 2, 2026, on Harvey's blog), co-founder Gabe Pereyra argues that the legal industry is the next sector to be reorganized by AI agents. He frames the transformation through intelligence replacing hierarchy, a framing originated by Sequoia Capital that he extends.

Pereyra's central claims:

• Engineering reorganized first; legal is next. Software lives inside a machine-readable loop, and labs prioritized coding capability because code is how the next generation of systems gets built. Legal follows, he argues, because it can be similarly structured around matters treated as standalone world models.
• Each legal matter as a standalone world model. Each matter and its associated documents, messages, research, and workflows can be analogized to a standalone world model within which teams of AI agents operate.
• The coordination bottleneck shifts. Pereyra writes that Harvey's engineers are now so productive they are harder to coordinate, with bottlenecks shifting from implementation to review, prioritization, coordination, and operating design.
• Legal's dual role. Pereyra argues in-house legal teams will both be transformed in their direct work and serve as stewards for AI implementation across their companies, increasingly asked to govern how the rest of the company uses agents, including where accountability sits, what risks are tolerable, and how trust is earned.

To illustrate his thesis that even insiders are still being blindsided by recent capability gains, Pereyra recounts "agent-pilling" his retired parents, both PhD computer scientists (including a former Apple autocorrect lead and a Stanford CS professor), by using Claude Code to add comprehensive test coverage to an open-source scientific computing library in 20 minutes.

Legal Agent Benchmark (LAB)

Harvey open-sourced its Legal Agent Benchmark (LAB) on May 6, 2026, with more than 1,200 tasks across 24 practice areas and over 75,000 rubric checks. The benchmark uses a strict "all-pass" grading rule: a deliverable counts only if every expert-written rubric criterion is met.

Harvey reported initial results on June 1, 2026. Under the all-pass rule, leading frontier models passed under 10% of partner-level legal tasks. Harvey described the pattern as "jagged intelligence," with no single model dominating: ChatGPT-5.5 led on regulated-industry and research work, while Claude Opus 4.7 led on corporate transactions. The results showed a steep cost-latency trade-off, with Claude Opus 4.7 costing roughly $50.90 and about 22 minutes per task, while a faster Gemini 3.5 Flash was described as nearly unusable for high-precision legal work. (Sources: harvey.ai; aigovernancelead.substack.com) See Legal Services — AI Deployment, AI Coding Agents.

Relationships

• deploys-in: Legal Services — AI Deployment
• competes-with: Legora, Thomson Reuters CoCounsel, Springbok AI (Cleary in-house)
• depends-on: Anthropic / OpenAI (frontier model providers); Claude Opus 4.5, Claude Opus 4.6, GPT-5.3 Codex cited as the underlying capability substrate
• introduces: Intelligence Replaces Hierarchy
• related: Agentic AI, Anthropic Economic Index — March 2026: Learning Curves (legal among top Claude-use categories)

Sources

• Fortune / $100m AI startup that wants to kill the billable hour (2025-09-03)
• Law360 Pulse / Clients Are Driving The AI Revolution In Law Firms (2025-10-22)
• How Autonomous Agents Will Transform Legal (Pereyra / Harvey, April 2026) (2026-04-02) — Pereyra "Legal Is Next" essay
• Harvey company website (Professional Class AI) — used as source for company description; not a separate sources page