Harvey, which sells AI tooling to law firms and in-house legal teams, says it is using OpenAI's GPT-6 Astra to pull more matter context into drafting and produce more structured legal documents. In a customer story on OpenAI's site, cofounder and president Gabe Pereyra frames the change in terms of input volume rather than raw model speed.
"We can give more context to the model and produce better and better structured outputs," Pereyra said in the post.
The claim is a workflow claim, not a benchmark claim. Harvey says the model supports lawyers as they work through the material that shapes a matter — reading, organizing, and drafting from court records, firm documents, case law research, and other legal reference material. Compared with other models, Harvey reports substantial improvements in document formatting and context awareness, which it says helps customers get more complete documents that better reflect the underlying material.
No scores, evaluation tables, or pricing appear in the disclosure, and the two companies have not published an independent comparison. That makes this a deployment and integration story rather than a product launch with its own numbers.
What's new
The concrete addition described in the post is a memory panel that carries an individual lawyer's preferences into the drafting workflow. Lawyers can record formatting and sourcing choices — for example, a preference for numbered lists, treating EDGAR as a priority source, or marking issues by priority through color.
The panel displays those preferences next to the source material and the draft memorandum, so the lawyer can see at a glance how the output is being guided. Harvey says that because GPT-6 Astra can process more context, it can produce higher-quality legal documents while customers spend more of their time on strategy.
In practice, the panel is a control surface: it sits next to the draft rather than replacing it, and it lets a lawyer steer formatting and sourcing without rewriting the prompt each time. The post does not describe how preferences are stored, whether they are shared across a team, or how they interact with firm-level policy.
Why it matters
Legal drafting is one of the few enterprise AI use cases where the output is a document a client may file or rely on, so formatting and source fidelity matter as much as reasoning quality. A model that can hold more of a matter in context reduces the number of times a lawyer has to re-supply background, which is where much of the friction in current legal AI workflows sits.
The memory panel also points at a broader pattern: personalization is moving from prompt engineering into product surface. If preferences live in the tool rather than in a saved prompt, the vendor controls how portable that configuration is across models — a detail that matters to firms that want to avoid lock-in.
Our take
The interesting part of this disclosure is what it does not contain. Harvey has a strong incentive to publish numbers, and it published none — no accuracy figures, no comparison table, no client results. That is not unusual for an integration note, but it leaves the formatting and context-awareness gains vendor-reported and untested outside Harvey's own workflows. Buyers should treat the memory panel as the more durable claim, since preference handling is a product decision rather than a model property.
