# OpenAI Puts GPT-6 Astra Behind a Legal Research Product Aimed at Large Law Firms

Astra for Law answered 54% of 200 questions from Vals AI's private Legal Research Bench set correctly, against 38.7% for GPT-6 Astra with web search alone.

- Published: 2026-09-18T06:27:34.772Z
- Canonical: https://polylog.news/ai/2026-09-18/openai-puts-gpt-6-astra-behind-a-legal-research-product-aime
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [OpenAI](https://openai.com/index/astra-for-law), [Polylog editors](https://polylog.news), [OpenAI (Cooley case study)](https://openai.com/index/cooley-gopublic)

OpenAI released [Astra for Law](https://openai.com/index/astra-for-law), a configuration of its GPT-6 Astra model built with a legal search layer, firm-level workflow tools and data controls designed for confidential client work. The search index covers more than 230 million uniform resource locators of United States case law, statutes, regulations, court rules and administrative decisions, [according to the launch materials](https://t.me/aipost/8179) and [independent coverage of the release](https://siliconangle.com/2026/09/17/openai-launches-astra-for-law-a-gpt-6-configuration-for-legal-research/).

The number that matters is the one OpenAI chose to publish. On 200 questions drawn from the private validation set of Vals AI's Legal Research Bench, Astra for Law answered 54 percent correctly, compared with 38.7 percent for GPT-6 Astra using ordinary web search. That is a large relative gain from retrieval grounding on a hard task. It is also a system that gets close to half of these questions wrong, on a benchmark selected and reported by the vendor, without third-party replication. Engineers evaluating it should treat the gap as a measure of what a curated legal corpus buys, not as a competence threshold for unsupervised legal work.

Distribution is the more consequential part of the announcement. OpenAI is offering the product first to selected firms through Trusted Access inside ChatGPT and Codex. An application programming interface (API) release is planned as gpt-6-astra-law, along with zero data retention for eligible firms and 26 ecosystem plugins connecting ChatGPT to legal tools including Relativity and Clio, [as LawSites reported](https://www.lawnext.com/2026/09/openai-releases-astra-for-law-a-gpt-6-model-configured-for-legal-work.html). Sullivan & Cromwell, Ropes & Gray and Cooley have already built custom tools on ChatGPT Enterprise with OpenAI's forward-deployed engineers. Cooley's [GO Public workflow](https://openai.com/index/cooley-gopublic) applies the model to initial public offering preparation, surfacing issues earlier in the document review cycle.

The competitive threat targets the legal artificial intelligence middle layer directly. Harvey, Thomson Reuters and LexisNexis have sold the combination of a frontier model, a proprietary legal corpus and firm-level controls. OpenAI now sells that same bundle itself, with the model vendor's cost advantage and a plugin surface that treats incumbent legal software as a connector rather than a platform.

## What this means

The exposed parties are the legal technology vendors whose product was integration and retrieval rather than the underlying model, because OpenAI just moved both inside its own stack and priced them as features. Law firms gain a cheaper path to internal tooling, and the constraint shifts from model access to data governance and malpractice risk, which is where zero data retention and audit controls become the real product. For model vendors, vertical bundles like this are how a frontier lab converts capability into per-seat revenue that does not erode when the next price cut arrives in the general API.

## What to watch

- Whether an independent evaluation reproduces the 54 percent figure on the Legal Research Bench, since a vendor-run private-set score is the weakest form of benchmark evidence.
- How Harvey, Thomson Reuters and LexisNexis respond on price and on exclusive data rights, which will show whether proprietary legal corpora still command a premium.
- Whether large firms bill clients for model-assisted work or absorb the cost, because that choice determines how quickly the tooling spreads beyond pilot groups.