Harvey GPT-6 Astra: Context-Aware Legal Drafting

Harvey GPT-6 Astra is a customer deployment that brings matter-specific sources and lawyer preferences into a legal drafting workflow. OpenAI says Harvey uses the model to analyze, synthesize and draft from court information, firm documents, case-law research and other legal context. Harvey reports stronger document formatting and context awareness compared with other models. Read OpenAI’s customer story.

Harvey GPT-6 Astra: how the drafting workflow works

OpenAI shows three visible parts of the workflow: the material behind a matter, a draft memorandum and the lawyer’s individual preferences. The material can include court information, internal firm documents and case-law research. The preference layer can tell the system to use numbered lists, prioritize EDGAR as a source or color-code issues by priority.

Those preferences appear alongside the source material and the draft in Harvey’s memory panel. That arrangement gives a lawyer a visible place to guide how the document is organized and which sources receive attention. OpenAI says GPT-6 Astra can process more context, allowing Harvey to produce more structured documents that better reflect the material supplied to the workflow.

The practical unit to evaluate is Harvey’s assembled workflow rather than the base model alone. The announcement says Harvey supplies legal context, but it does not describe the retrieval method, context-selection rules, document limits or permission checks used for each source. Those implementation choices determine which material reaches the model and which instructions govern the result.

Our GPT-6 Astra model page covers the model’s general capabilities and current API facts. The Harvey story adds a specific use case: turning a controlled set of legal materials and individual drafting preferences into a document inside a legal-work platform.

What improved in Harvey’s legal drafts

Harvey says it observed substantial improvements in document formatting and context awareness with GPT-6 Astra when compared with other models. OpenAI describes the resulting documents as more complete and more reflective of the underlying material. The release does not attach a score, task count, acceptance threshold, named comparison model or before-and-after sample, so the result cannot be converted into a public error rate or productivity figure.

Formatting matters in legal work because structure carries meaning. Numbered sections, issue labels and consistent source treatment can make a memorandum easier to inspect and revise. Context awareness matters for a different reason: a polished draft is not useful if it ignores a controlling fact, mixes records or imports an unsupported assumption. A pilot can measure both qualities by checking issue coverage, source support, required structure and the corrections needed before acceptance.

This event is separate from the earlier Astra for Law launch. Astra for Law concerns a restricted OpenAI configuration with a dedicated legal search index and specific access terms. The Harvey announcement concerns a customer product using GPT-6 Astra in Harvey’s own drafting environment. A reader should not infer Astra for Law access, its planned API identifier or its legal-search benchmark from this customer story.

How individual lawyer preferences change the output

Harvey’s examples show three different kinds of control. A request to use numbered lists changes presentation. A request to prioritize EDGAR changes source selection. Color-coding issues by priority adds a review signal to the draft. Testing each rule separately makes it possible to see whether a failure came from formatting, source choice or issue classification.

For presentation rules, reviewers can check whether the structure persists after edits and across longer documents. For source priorities, they can verify that the preferred source supports the proposition and that contrary material remains visible. For issue labels, they can change one priority and confirm that only the intended passages move or change color. This is our proposed evaluation method, not a test result published by Harvey or OpenAI.

The customer story identifies the settings as individual preferences but leaves their operating scope open. It does not say whether a personal rule can be overridden for one matter, whether firm templates take precedence, how long a preference persists or who can inspect its history. A useful control panel should make the active instruction, owner and scope visible before generation. Otherwise, a correct personal convention could be applied to the wrong client or document type.

Source access deserves a parallel check. Before scoring draft quality, a firm should confirm which repositories the workflow can reach, whether access follows the current user’s permissions and where generated documents are saved. Our GPT-6 API overview explains the general developer path, but it does not describe Harvey’s private configuration or commercial availability.

An editorial test for Harvey-style preference memory

The following is an untested editorial example designed around Harvey’s distinctive preference controls. Create one approved source packet and one memorandum template, then define three visible rules: use numbered issues, prefer EDGAR for company filings and color-code unresolved questions by priority. Generate the same short memorandum under three profiles: no saved preferences, the lawyer’s saved profile and a one-matter override.

Compare the outputs with a small matrix. Record whether each rule activated, whether the selected EDGAR material actually supports the statement, whether priority colors match a prepared issue list and whether the override stayed confined to the intended matter. Then revise one fact and one preference. The new draft should preserve unaffected sections, update the targeted passages and show which rule produced the change.

This test isolates the product claim more directly than grading a single polished document. It asks whether individual instructions remain predictable when context changes and whether a matter-specific exception can be distinguished from a saved personal convention. It also produces evidence a reviewer can inspect: the active profile, source set, draft differences and accepted corrections.

Teams comparing models can use our model comparison tool for published model-level differences, then evaluate Harvey’s full workflow with their own documents, permissions and review standards. The verified development is that Harvey has integrated GPT-6 Astra into legal drafting and reports better structure and context awareness. The customer story supports that workflow change; it does not quantify accuracy or time saved.