Harvey Leans on GPT-6 Astra for Longer-Context Legal Drafting

OpenAI has published a customer story describing how Harvey, a legal AI company serving law firms and in-house teams, is using GPT-6 Astra to draft legal documents that pull in more of the surrounding context of a matter. According to OpenAI, Harvey has seen "substantial improvements" in document formatting and context awareness versus other models, and the company is pairing that with a memory panel that lets individual lawyers encode drafting preferences. No benchmarks, accuracy figures, customer names, pricing or deployment scale were disclosed.

What the post actually claims

The story, dated September 23, 2026, is part of a run of OpenAI startup spotlights — others in the same series cover Parallel, Higgsfield AI and V7 — and it follows the familiar format: a named executive quote, a description of the workflow, and a directional claim about quality. OpenAI lists the product in use as its API, and tags Harvey as a North America-based startup in the technology sector.

The substance is narrow but specific. OpenAI says Harvey's customers use the platform to turn large volumes of information into complex legal documents across workflows that range from litigation to mergers. With GPT-6 Astra, Harvey says it can push more of that material into the drafting process at once — court information, the firm's own documents, case law research and other sources of legal context — and get back outputs that are more structured and more complete.

"We can give more context to the model and produce better and better structured outputs," said Gabe Pereyra, Harvey's cofounder and president, in the quote OpenAI published.

That is a quality claim stated in relative terms. OpenAI does not say what Harvey measured, against which models, on what document types, or how much better "substantial" means. For a domain where a misstated citation or a dropped defined term is a professional liability issue, the absence of any evaluation detail is the most notable gap in the announcement.

The memory panel is the more interesting product idea

The second half of the story describes Harvey's memory panel, which brings an individual lawyer's preferences directly into the drafting workflow. OpenAI gives three examples: using numbered lists, prioritizing EDGAR as a source, and color coding issues by priority. Those preferences sit alongside the source material and the draft memorandum, which OpenAI frames as giving the lawyer a clearer way to steer the output.

This is a quieter but more durable idea than raw context length. Legal drafting is full of house style — how a firm numbers recitals, which sources it trusts, how it flags risk — and most of that knowledge lives in a partner's head or in a template nobody has updated since the last deal. Turning it into explicit, visible, editable instructions does two things at once: it improves the first draft, and it makes the model's behavior auditable. A lawyer can look at why an output came out a certain way and change the rule rather than re-prompting.

It also sidesteps a common failure mode of implicit memory systems, where a model silently accumulates preferences and users lose track of what is influencing generations. Putting the preferences in a panel next to the draft is a design choice in favor of legibility. OpenAI does not say whether preferences can be shared across a team or a matter, whether they are enforced deterministically or simply passed as instructions to the model, or how conflicts between preferences get resolved. Those details matter a lot in practice and were not disclosed.

Longer context is necessary, not sufficient

The throughline of the post — more context in, better documents out — is the central bet of most serious document-heavy AI products right now. It is also the claim that deserves the most scrutiny.

The research community has spent the past several model generations documenting that a model's ability to accept a large context is not the same as its ability to use all of it reliably. Retrieval quality degrades unevenly across long inputs, and stuffing more source material into a prompt can dilute attention on the passages that actually decide a question. OpenAI's post does not disclose GPT-6 Astra's context window, how Harvey chunks or ranks the material it sends, or whether Harvey still uses retrieval to pre-filter. "Can process more context" is the only technical characterization given.

For legal work specifically, the bar is not "did the draft mention the right case" but "did it characterize the holding correctly, in the right jurisdiction, at the right procedural posture." Formatting improvements — which OpenAI explicitly calls out — are real and useful, since a well-structured memo is faster to review. But formatting is also the most visible and least consequential axis of quality. A cleanly numbered document with a hallucinated citation is worse than a messy one, because it reads as more trustworthy.

Nothing in the post suggests Harvey is unaware of this; the company's pitch has long centered on keeping lawyers in the loop, and OpenAI's framing explicitly positions the output as a draft that a lawyer guides and reviews. But readers evaluating the claim should note that the announcement is a workflow description, not an evaluation.

What this signals about the vendor layer

The broader pattern worth watching is how much differentiation remains in the application layer as frontier models absorb more capability. Harvey's answer, as presented here, is context assembly plus preference management plus secure deployment into firm environments — the connective tissue between a model and a regulated workflow, rather than the reasoning itself.

That is a defensible position, but it is also one that gets renegotiated with every model release. Each generation that handles more context natively removes some of the engineering that vendors previously sold as their edge. The vendors that keep their footing tend to be the ones whose value sits in data access, permissioning, audit trails and domain-specific evaluation — things a general-purpose API does not provide. OpenAI's post touches on the first of those and is silent on the rest.

For product teams outside legal, the transferable lesson is the memory panel pattern: make the implicit rules of your domain explicit, visible and user-editable, and place them where the user is already looking. That is a design decision, not a model capability, and it does not require waiting for the next release.

What to watch

• Whether Harvey or OpenAI publishes any evaluation data — citation accuracy, hallucination rates, reviewer edit distance — to support the claim of substantial improvement over other models. None was provided.
• How the memory panel scales beyond the individual: whether preferences can be set at the firm, practice-group or matter level, and how conflicting rules are resolved. Not addressed in the announcement.
• Whether context volume translates into reliability on long documents, or whether retrieval and ranking still do the decisive work behind the scenes.
• How law firms and their clients treat AI-drafted work product under confidentiality and privilege obligations as more matter context flows into third-party model APIs.
• Whether OpenAI's startup spotlight cadence continues to favor directional customer quotes over measured results, which would tell you something about what the model can currently be claimed to do in public.