Harvey just told the world it is building its own legal foundation model. Not a fine-tune. A whole new model series, reportedly in development as of June 18, 2026. The same company that started in 2022 and was reportedly a wrapper on top of GPT.
Here is the proof that this matters. Harvey charges around $1,200 per seat per month. Direct access to a frontier model like Claude.ai Team runs about $25 per seat per month. That is a 48x price gap. Customers are paying it.
So why would a company sitting on an $11 billion valuation, built entirely on rented intelligence, decide to start training its own brain? That is the question. And the answer tells you where every vertical AI company is heading next.
The Ownership Ladder
Here is the framework I want you to remember: The Ownership Ladder.
The math behind owning the model instead of renting it.
Every vertical AI company starts on the bottom rung. They rent intelligence. They call an API, slap a UI on top, and sell the wrapper. Cheap to start. Easy to copy. Zero moat.
The next rung up is orchestration. You stop betting on one model and route across many. Harvey reached this rung by May 2025, routing across OpenAI, Anthropic, and Google. One prompt can trigger up to 100 model calls. Each model gets benchmarked by Harvey's own lawyers.
The third rung is owning your evaluation layer. Harvey built BigLaw Bench, a legal benchmark scored by real attorneys. Harvey says its proprietary models outperform leading foundation models on that benchmark. That benchmark is an asset competitors cannot see.
The top rung is owning the model itself. That is the rung Harvey just stepped onto. The ladder is simple: rent, route, evaluate, own. Each rung deepens the moat. Each rung makes you harder to rip out.
Why Owning the Stack Beats Renting It Forever
Let me pull back and think about this the way a long-arc strategist would. The question is not "can Harvey build a model." The question is "what kind of asset are they building, and does it compound?"
Start with the math. If 98% of Harvey's seat price sits above the raw model cost, then the model is Harvey's cost of goods sold. Every dollar paid to OpenAI is a dollar Harvey does not keep. Owning the model turns a recurring expense into a one-time capital investment. Only cash is real. The rest is accounting.
Now think about leverage. When you rent your intelligence, the landlord can move down the stack and become your competitor. General-purpose labs are already pushing into legal directly. A renter has no defense against the company that owns the building. He knows exactly how thin a wrapper's protection really is.
This is counterpositioning in its purest form. Harvey is not trying to out-reason GPT-5 on every task. Pereyra was explicit: the new model series will work as an agentic system that learns to control legal tools and is intended to escalate to frontier models when needed. Harvey owns the narrow domain and rents the broad reasoning. That is a deliberate split.
Here is the strategic line that matters most. Pereyra's stated second goal is to "create the foundations for law firms to build their own specialized models and own their own intelligence." Read that again. Harvey is not just owning its own stack. It is selling firms the tools to own theirs.
That is the flywheel. Harvey runs proof-of-concepts with firms to train open-source models on their playbooks and client patterns. Digital twins of how a firm actually works. Once a firm encodes its judgment inside Harvey, switching out is no longer changing an API key. It is abandoning years of institutional memory.
Now the honest part. It is unclear whether the model moat runs as deep as the workflow moat. A lot of legal work runs on public case law and standard clauses. Skeptics in legal-tech commentary argue the real defensibility lives in workflow design and distribution, not in proprietary base weights. I think they are partly right.
My read is that Harvey knows this too. The model is not the moat by itself. The moat is the full system: model, plus firm data, plus benchmarks, plus agentic workflows, fused so tightly that no single piece can be cleanly extracted. The base weights matter less than the assembly.
The risk is real. Training and serving your own models is expensive. Frontier models keep getting cheaper and smarter. If the gap closes faster than Harvey can specialize, the capital spent on owning could become a stranded asset. The data on whether post-trained open-weight models hold up against frontier systems is still mixed. Harvey says early experiments with partners including Fireworks AI and Nvidia show post-trained open-weight models could approach frontier performance on legal tasks. The full numbers are not published.
So this is an asymmetric bet, not a sure thing. The downside is bounded by Harvey's cash. The upside is owning the intelligence layer of a profession that pays six figures per deployment. That is the kind of bet a strategist takes.
2031
Three signals inside the same shift
Renting intelligence is now the cost of goods sold.
Harvey charges $1,200 a seat against roughly $25 for raw frontier access. With 98% of the price sitting above model cost, every dollar paid to a lab is a dollar not kept. Owning the model turns a recurring expense into a one-time capital bet.
The landlord can become the competitor.
General-purpose labs are pushing into legal directly. Harvey reached the orchestration rung by May 2025, routing up to 100 model calls per prompt across OpenAI, Anthropic, and Google. Owning the narrow domain while renting broad reasoning is a deliberate defensive split.
Winners own exactly one layer.
By 2031 the vertical market likely splits into thin rented wrappers and stack owners that look like infrastructure. The durable pattern is the hybrid middle: own the narrow high-stakes parts, keep renting frontier reasoning for the rest.
Pull the lens out five years. By 2031, I expect the phrase "AI startup" to sound as dated as "internet company" does now. The interesting question will not be whether a company uses AI. It will be how far up the Ownership Ladder it climbed.
The vertical AI market will likely split into two camps. One camp stays on the rented rung. Thin wrappers, low margins, easy to disrupt the moment a lab ships a competing feature. The other camp owns its stack, its data, and its evaluation layer. These companies will look less like software vendors and more like infrastructure.
Harvey is betting that law belongs in the second camp. The logic extends to health, finance, and any field where hallucination costs are measured in malpractice exposure, not annoyance. In high-stakes verticals, trust is the product. You cannot fully outsource the thing your customers trust you for.
I think the durable pattern by 2031 looks like the hybrid middle. Vertical companies will own the narrow, proprietary, high-stakes parts. They will keep renting frontier reasoning for the broad, general parts. Owning everything is too expensive. Owning nothing is too dangerous. The winners will own exactly the layer their customers cannot live without.
What to Build This Weekend
You do not need a research lab to climb the first few rungs. You need to stop renting blindly and start owning your evaluation layer. That is the cheapest moat available to anyone.
First, pick one narrow task in a domain you understand. Not "legal AI," something tiny. Contract clause classification. Product description quality. One task you can judge yourself.
Second, build your own mini-benchmark. Write 20 example inputs and the correct outputs by hand. This is your BigLaw Bench in miniature. Run three different models against it: one from OpenAI, one from Anthropic, one open-weight. Score them yourself. This is orchestration plus evaluation, the second and third rungs of the ladder, built in an afternoon.
Third, automate the boring part around it. If you sell physical products, a tool like Product Upload copies listings from supplier websites and pulls many product specifics, so your model has clean inputs to work on. If you are building something travel-related, Pop-Plan uses an AI assistant and a quiz to generate itineraries. Study how it wraps the model in a workflow, because the workflow is where the value hides.
A quick definition. An evaluation layer just means a repeatable test that scores model outputs against answers you trust. No CS degree required. A spreadsheet and 20 honest judgments will do.
Expect your first models to fail your own benchmark. Good. That failure is the data. The point is not to ship a perfect system this weekend. The point is to own the test that tells you what "good" means in your domain. Get your reps in. The companies winning in 2031 started by owning that one small thing.
Stop renting blindly and start owning your evaluation layer.
- Pick one narrow task. Not "legal AI," something tiny like contract clause classification or product description quality. One task you can judge yourself in a domain you actually understand.
- Build a mini-benchmark. Write 20 example inputs and the correct outputs by hand, then run three models against it: one from OpenAI, one from Anthropic, one open-weight. Score them yourself. This is your BigLaw Bench in miniature.
- Automate the boring part around it. Wrap the model in a repeatable workflow that feeds it clean inputs, because the workflow is where the value hides. Expect your first models to fail your own benchmark; that failure is the data.
The model is not the moat. The assembly is.
Harvey is not betting it can out-reason a frontier model on every task. It is fusing model, firm data, benchmarks, and agentic workflows so tightly that no single piece can be cleanly extracted. The risk is real: training and serving your own models is expensive, and if frontier systems get cheaper faster than Harvey can specialize, the spend becomes a stranded asset. But the downside is bounded by cash while the upside is owning the intelligence layer of a profession that pays six figures per deployment. That is an asymmetric bet, and the question by 2031 will not be whether a company uses AI but how far up the Ownership Ladder it climbed.