Astra for Law: Access, Legal Research and Limits
OpenAI announces Astra for Law for selected law firms, with a dedicated legal search index, tailored instructions and an API still coming soon.
Astra for Law is OpenAI’s new configuration of GPT-6 Astra for professional legal work, announced on September 17, 2026. It combines the model with a dedicated legal search index and tailored instructions. Initial access is restricted, and the API is still described as coming soon. Read the announcement.
What Astra for Law changes for legal research
The Legal Search Index covers U.S. case law, statutes, regulations, court rules and administrative decisions, with new sources added daily. OpenAI’s help documentation describes tasks such as researching a client’s facts, developing arguments or deal terms, and examining objections. It also tells users to review answers and cited sources before relying on them. Official product guidance.
OpenAI reports 54.0% overall correctness for the complete configuration on 200 U.S. questions from a private Legal Research Bench validation set, versus 38.7% for GPT-6 Astra with web search alone. Both used their highest reasoning effort. The reported 40% improvement is relative, not a 40-percentage-point gain. These are vendor-reported benchmark results, not our testing. Evaluation details.
Our interpretation is that a legal team should evaluate the whole research workflow. A useful answer needs evidence that a reviewer can inspect, and a polished paragraph alone is an incomplete acceptance criterion. The benchmark also leaves substantial room for errors. It does not establish reliability on a firm’s own matters, documents or jurisdictions.
Who can get GPT-6 Astra Law access?
The help page specifies selected U.S. law firms, with access for eligible lawyers and people working under their supervision. To request early access, contact an OpenAI account team or OpenAI Sales. The model-picker label is GPT-6 Astra Law. The announcement places initial access in ChatGPT and Codex through Trusted Access; its planned API identifier is gpt-6-astra-law. Access guidance.
For a firm planning a pilot, our recommendation is to establish the account and workspace that will receive access before assigning work. Record who will supervise the pilot and which task it should demonstrate. Do not schedule an API integration around a model name alone while availability remains forthcoming.
Our general GPT-6 access guide explains the broader product entry points. This launch adds a separate eligibility question for legal work; access to a general model does not establish admission to this program.
How legal plugins fit into a pilot
The Astra help page gives two concrete examples: iManage can support drafting a negotiation brief and saving it to a matter file, while DeepJudge can bring earlier deals into a comparison. These describe possible workflows; each plugin’s listing determines its access requirements. Legal plugin examples.
OpenAI’s plugin documentation separates installing a package from authorizing its connected services. A plugin can contain skills, apps or setup templates, and some contain only instructions. Availability varies by account, workspace, role, region and product surface. Installing one does not grant new rights to the underlying files. Plugin setup documentation.
For a supported plugin, the documented starting sequence is to open Plugins, inspect the included apps, install the plugin, and complete any required connection flow. An eligible administrator must configure supported sync. In a supported Codex task view, use Sources, then Use plugins, to select an installed plugin.
Our suggested pilot check is simple: have an authorized reviewer retrieve one approved document and identify its source. Then test the intended output separately, such as a comparison or a proposed draft. If the workflow writes to a matter system, inspect the destination and resulting file. Record an unavailable connection as a setup problem rather than treating it as a model-quality result.
What the privacy commitments cover
For eligible firms, OpenAI’s announcement includes API Zero Data Retention and excludes ChatGPT Enterprise usage from human review by default. These are program-specific commitments. Launch privacy statement.
The broader enterprise privacy policy says business inputs and outputs are not used for model training by default. Enterprise workspace administrators control retention and which internal sources connect. The API’s general retention policy can differ: eligible endpoints and qualifying use cases can receive ZDR, while other API inputs and outputs may be retained for up to 30 days, subject to stated exceptions. Enterprise privacy policy.
Our practical recommendation is to ask the account team which commitment applies to each proposed component: the legal model, the ChatGPT workspace, the API and each connected provider. Put those answers next to the pilot’s actual data flow. A firm can then assess the planned configuration with its own requirements. Our Enterprise administration update provides related background on investigating workspace access.
An example evaluation before live matter work
The following is an editorial example, not a tested prompt or an OpenAI product guarantee. Begin with a fictional contract dispute and an explicitly named U.S. jurisdiction. Ask for a short research memo that separates the supplied facts, unresolved questions, supporting authorities and the strongest opposing argument.
A reviewer can score the result in three passes. First, open every cited authority and check whether the stated proposition appears in it. Second, assess whether the answer distinguishes binding authority from other useful material and acknowledges contrary evidence. Third, compare the draft with a reviewer-prepared checklist of issues, recording omissions as well as incorrect statements.
Keep the original prompt, output and review notes together. Repeat the same task after changing one factual assumption, and inspect whether the answer explains why its conclusion changed. This produces a concrete acceptance record for that task without treating a single convincing response as proof of broad competence.
The immediate decision for a law firm is whether this restricted launch merits a controlled evaluation. API builders should confirm availability directly before starting implementation; our GPT-6 API overview covers general API concepts, while this announcement establishes the separate legal offering.