Harvey just declared independence from its AI landlords — and every other vertical SaaS company paying per-token rent is watching.
The Summary
- Harvey launched Tenet, its first proprietary legal AI model, designed to handle multi-hour legal tasks at lower cost than the third-party models (OpenAI, Anthropic) it currently relies on
- The move is defensive: Anthropic and OpenAI are both pushing directly into legal workflows, threatening to disintermediate Harvey's $11B business
- Custom models mean Harvey controls costs, owns its differentiation, and can eventually let law firms train models on their proprietary data
The Signal
Harvey built an $11 billion legal-software business as a wrapper. It took the best foundation models from OpenAI and Anthropic, added legal-specific prompting and workflows, and sold access to law firms. Smart arbitrage. But arbitrage windows close.
Now the model providers want Harvey's customers. Anthropic is building legal plugins for document review. OpenAI hired Ironclad's founder to lead its legal push. Google and Meta are circling. Harvey's suppliers became its competitors.
"What happens when your supplier decides it wants your customers, too?"
The economics forced the move. Every time a lawyer uses Harvey, the company pays OpenAI or Anthropic for the inference call. High usage means high costs. Great for growth metrics, terrible for margins. Harvey cofounder Gabe Pereyra, a former Google DeepMind researcher, saw the math: build a model that's good enough for 60-80% of legal tasks, route that work through Tenet instead of paying per-token to OpenAI, and margins improve without raising prices.
But cost control is table stakes. The real signal is ownership. A proprietary model means Harvey can train on law firm data without sharing it with OpenAI. That unlocks the next business model: custom models per firm. A Magic Circle firm doesn't want its precedents, strategies, and client patterns feeding into a shared model every competitor uses. They'll pay for isolation. Harvey can now offer it.
Key implications:
- Harvey moves from model aggregator to model builder, owning its differentiation
- Per-firm custom models become the premium tier, not just workflow automation
- Every vertical SaaS company using foundation models faces the same build-or-rent calculus
This is the agent economy's first major verticalization play. Foundation models are commoditizing. The value migrates to domain-specific training, workflow integration, and trust. Harvey is betting that "legal AI" isn't a feature OpenAI can bolt on. It's a moat you build with years of case law, firm-specific training data, and lawyers who trust you not to leak their strategy to the other side.
The Implication
If you're building vertical AI software on top of OpenAI or Anthropic, you're on borrowed time. Your model provider is either coming for your customers or will let a competitor undercut you. Harvey's move is the playbook: build a good-enough model for your core use cases, own the training data pipeline, and sell customization as the premium product.
For law firms, this changes the buying decision. You're no longer choosing between Harvey and doing legal work manually. You're choosing between a shared model anyone can access and a private model trained on your firm's decades of work. The second option costs more. It's also the only one that compounds your advantage.