Harvey pivots to multi-model workflows

This marks Harvey’s shift from being a legal model company to being a legal workflow company. Once frontier models beat its in house system on Harvey’s own legal benchmark, the scarce asset stopped being raw reasoning quality and became everything around it, routing the right model, pulling the right documents, enforcing firm permissions, and packaging the output into repeatable work like diligence, review, drafting, and research.1 2 3 4

• Harvey’s mid 2025 update says Google Gemini 2.5 Pro, xAI Grok 3, OpenAI o3, and Anthropic Sonnet 3.7 all outperformed Harvey’s fine tuned model on BigLaw Bench. That is why the company moved to multi model agentic workflows instead of defending a proprietary core model.1 2
• Large firms do not buy these tools like normal SaaS. Interviews show they start with 5 to 20 seats, hot swap licenses, and judge success by completed matters and daily usage. In that setup, vendor advantage comes from onboarding, workflow fit, and integration, not owning a special model.3 4
• This also narrows the gap with rivals. Legora skipped fine tuning and focused on workflow UX and international fit. Specialists like Spellbook, DeepJudge, and Luminance are attacking narrower jobs, while incumbents like Thomson Reuters and Clio are bundling AI with research and system of record products they already own.3 4 5

The next phase of legal AI will be won by the products that turn foundation models into trusted firm infrastructure. Harvey is heading toward a model router plus agent layer that sits on top of law firm systems, while the market around it keeps splitting between broad workflow platforms, specialist tools, and incumbents with proprietary legal content and distribution.2 3 4 5