Mostly because of one question clients ask: where do our documents go? Harvey runs on Claude, GPT and Gemini, the same models any firm can subscribe to, so the model is not the reason to pay more. The reason is a reply to that question that the firm can sign, which legal AI platforms sell with a seat contract attached. A firm that can give the reply itself no longer has to buy it.
The model is not the difference
Harvey started on OpenAI's models and added Anthropic and Google in May 2025. In March 2026 it named Claude Opus 4.6, Sonnet 4.6, GPT-5.2 and Gemini 3 Flash among the models it uses. A firm paying for Harvey is paying, in part, for Claude and GPT.
On legal research, the general tools are close. In the Vals AI benchmark of 210 questions, ChatGPT scored 80%, inside the range of the three specialist tools. The largest research platforms did not take part, so this does not settle everything. It does show that the model no longer justifies the price.
The large AI companies now sell legal products of their own. Anthropic's Claude for Legal, launched in May 2026, even connects to Harvey.
So what is the premium for?
Some of it pays for real work: legal workflows, research content, integrations. A firm that needs those has a reason to buy.
Another part pays for an answer. When a client's security questionnaire asks where its documents go, the firm needs a reply it can sign. Harvey says every model provider it uses must accept zero data retention, no human review of customer data and no training on customer inputs. The firm attaches the vendor's terms and the questionnaire is done.
This is what makes general tools feel like a hassle. Opening ChatGPT is easy. Explaining afterwards where the documents went is the hard part, and the platform does it for you, per seat, for at least a year.
Why the general tools leave the question open
OpenAI and Anthropic both say their business accounts are not used for training by default. That helps, but the client's documents still leave the firm.
In the United States, ABA Formal Opinion 512 says a lawyer needs the client's informed consent before entering information about the case into a self-learning AI tool. A line in the engagement letter is not enough. Every new tool reopens the same discussion with the client.
Answer the question yourself
Change what leaves. A model does not need to know that the tenant is Jane Doe at 14 Elm Street to review a lease clause. It needs the clause. Remove the names, addresses and account numbers first, and the reply to the client is short: they never left our machines.
That reply does not depend on any vendor's contract. It is true for Claude, for ChatGPT and for Harvey. The firm can pick each tool on price and features, and buy only the seats it uses.
With promptShield, identifiers become codes such as [P0ABCD] on your own computer. Within a folder, the same name keeps the same code, so the model can follow who is who across documents. Decode then puts the real values back on your machine. Detection and redaction run locally by default. Only the optional Deep Analysis feature can be set to send page text to a remote model. Details are in reversible redaction, explained.
What removing identifiers does not cover
It does not make a file non-confidential. ABA Model Rule 1.6 protects all information about the representation, not only names, and a merger memo with every name removed can still point to the target. So keep a business account that does not train on your inputs, and write the rule down: an AI policy for a small law firm covers how.
With that in place, a legal platform becomes a choice rather than a requirement. Buy it for its workflows, if they save more than they cost. Disclosure: we make promptShield, so we benefit when firms remove identifiers themselves.
Frequently asked questions
Is Harvey just ChatGPT for lawyers?
No. It uses models from Anthropic, OpenAI and Google and adds legal workflows, integrations and contract terms on data. The models themselves are available to any firm on a business account.
Can a law firm use ChatGPT or Claude on client work?
On a business account, with care. Your professional rules still apply, and ABA Formal Opinion 512 asks for informed consent before client information goes into a self-learning tool. Removing identifiers first means far less client information goes in.
Does removing names make a document non-confidential?
No. Rule 1.6 protects everything relating to the representation. Deal terms and distinctive facts can still point to the client.
Does Claude for Legal replace Harvey?
Not directly. It includes a connector to Harvey, so a firm can use both. The choice depends on workflows and research content, because the underlying models overlap.