Introducing Astra for Law

Legal work is high-stakes enough that “just use the chatbot” was never going to be enough. OpenAI is shipping a configured foundation: frontier model, legal search over case law, statutes, and regulations (including Free Law Project / CourtListener coverage of more than 99.9% of published U.S. precedential case law), and instructions aimed at analysis and writing.

On Vals AI’s private Legal Research Bench validation set (200 questions), Astra for Law hit 54.0% overall correctness at max reasoning versus 38.7% for GPT-6 Astra with web search alone—about a 40% relative lift—and found more reference cases and passages on audited sets.

Harvey, Legora, and firms like Sullivan & Cromwell are in the early orbit. Confidentiality controls and Latham-guided governance are the trust pitch.

We’re for tools that ground answers in authorities lawyers can check—and we’re not confusing a 54% bench pass rate with partner-ready autonomy.

Astra for Law shows up as GPT-6 Astra Law in the picker and gpt-6-astra-law in the API, initially via Trusted Access for selected firms, then API. Twenty-six ecosystem plugins connect to specialist tools firms already pay for. Zero Data Retention on API and ChatGPT Enterprise excluded from human review by default matter for client confidences; ethical walls and permissions work with Latham & Watkins is the governance story.

Useful AI in law looks like faster, better-sourced first passes that a human still owns. The downside if marketing outruns the bench: junior overreliance on answers that still fail nearly half of a research evaluation, or quiet leakage across matters despite ZDR promises. I’m watching real-matter win rates versus the Vals number—and whether bar and malpractice guidance catches up before deployment outruns oversight.

OpenAI announcement dated September 17, 2026. Combines GPT-6 Astra with a dedicated legal search index and firm-oriented Trusted Access controls.