Models

OpenAI Launches GPT-6 Astra for Law for Legal Work

OpenAI has launched Astra for Law, a specialized version of its GPT-6 Astra model designed to give legal professionals highly accurate research capabilities with strict privacy controls.

The Decoder4 days agoModels
Image: The Decoder

OpenAI has unveiled Astra for Law, a specialized version of its GPT-6 Astra model tailored specifically for legal research, analysis, and writing. The new system integrates the core model with a dedicated legal search index that spans more than 230 million URLs of United States case law, statutes, and regulations. To build this index, OpenAI sourced data from the Free Law Project, which provides coverage of over 99.9 percent of published US precedents.

In internal evaluations using Vals AI's Legal Research Bench, the specialized model demonstrated a significant performance leap over its general-purpose counterpart. Astra for Law successfully answered 54 percent of the 200 test questions. In contrast, the standard GPT-6 Astra model utilizing a basic web search achieved a pass rate of only 38.7 percent on the same benchmark.

To address the strict confidentiality requirements of the legal industry, OpenAI is offering law firms a Trusted Access program that features robust privacy controls, including a zero data retention policy. Additionally, the company is launching 26 plugins to connect the model with popular industry software such as Relativity and Clio. API access is also available, allowing legal technology developers like Harvey and Legora to build custom applications on top of the new model.

This release intensifies the competition in the legal AI space, where Anthropic is also actively expanding its footprint. For practitioners, Astra for Law offers a more reliable tool for drafting and discovery, reducing the risk of hallucinations by grounding the AI in a verified database of precedents. The combination of specialized search, API extensibility, and strict data privacy protocols makes it easier for firms to integrate advanced generative AI directly into their daily workflows.

This is our own summary of reporting by The Decoder

More in Models