OpenAI Built a Version of GPT-6 Astra That Only Does Law — and It Benchmarks 40% Better

OpenAI launched Astra for Law — a specialized deployment of GPT-6 Astra tuned specifically for legal research and analysis. The system combines the base model with a 230-million-URL legal search index covering US case law, statutes, and regulations drawn from the Free Law Project, which OpenAI claims covers over 99.9% of published US precedents.

The performance claim that defines the launch: Astra for Law achieves a 54% success rate on Vals AI’s Legal Research Bench (200 questions), compared to 38.7% for standard GPT-6 Astra with regular web search. That’s a 40% relative improvement by grounding the model in a curated, domain-specific corpus rather than the open web.

The Pattern

This is the fifth Stack Archive post covering GPT-6 Astra — following coverage of the AGI era announcement, ARC-AGI benchmark controversy, prompting patterns, and spatial reasoning capabilities. Each prior post addressed what Astra can do generally. Astra for Law is the first vertical-specific deployment: OpenAI packaging the model as an end-to-end legal research product rather than a general-purpose tool.

What Astra for Law Includes

The legal deployment is built on three components:

The legal index — 230M URLs from Free Law Project, covering US precedents back to early case law. The specificity matters: a model answering legal research questions against a curated index of actual court decisions performs substantially differently from one searching the open web, where legal information competes with consumer-facing summaries, law firm marketing pages, and outdated content.

Domain-tuned analysis and writing — tailored instructions for how the model approaches legal analysis, citation structures, and statutory interpretation. Legal writing has distinctive conventions (parallel citation formats, case reasoning chains, regulatory cross-references) that require model-level adaptation, not just corpus grounding.

26 legal-specific plugins — covering tools like Relativity (e-discovery platform) and Clio (law practice management software). Integration into existing legal workflows is what moves AI from research assistant to embedded tool. Relativity integration specifically signals that OpenAI is pursuing the discovery/litigation side of the market, not just research memos.

Privacy Architecture for Law Firms

Zero data retention for law firms in the Trusted Access program. This directly addresses the professional responsibility concern that has slowed law firm AI adoption: client confidentiality obligations under ABA Model Rules and state bar equivalents create liability when client matters are used to train external models.

Zero retention means Astra for Law processes queries without storing them in training datasets. For in-house counsel and Am Law 200 firms evaluating AI tools, this moves “data handling” off the risk list.

The Competitive Angle

Harvey and Legora — two of the more prominent legal AI companies — are API customers for Astra for Law. This is a meaningful signal: OpenAI is positioning itself as infrastructure provider for the legal AI market, not just a direct competitor to Harvey. If Harvey is built on Astra for Law, OpenAI captures value whether Harvey wins clients or a competitor does.

Anthropic is simultaneously expanding its legal AI offering, which the article notes creates “an emerging competitive market for specialized legal AI tools.” The vertical-specialization race — general models winning individual markets through domain-tuned deployments — is now clearly underway for regulated professional services.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.