OpenAI has introduced Astra for Law, a version of its GPT-6 Astra model configured for legal research and writing. It arrives carrying 230 million URLs of US case law, statutes, and regulations — which is, by most measures, more precedent than any human attorney has read, though attorneys will likely find a way to bill for the comparison.
The model passed 54 percent of the benchmark questions. The humans it is replacing average somewhat lower, though they do have bar licenses.
What happened
Astra for Law pairs the GPT-6 Astra model with a dedicated legal search index built on data from the Free Law Project, which covers over 99.9 percent of published US precedents. OpenAI ran its own benchmark using Vals AI's Legal Research Bench, where Astra for Law answered 54 percent of 200 questions correctly, compared to 38.7 percent for GPT-6 Astra using plain web search.
The improvement is real. Whether 54 percent is a passing grade is, appropriately enough, a matter of interpretation.
API customers including Harvey and Legora can build on the model. Law firms gain access through a Trusted Access program with privacy controls and zero data retention — a feature that, in a profession built on confidentiality, the humans will find reassuring rather than ironic.
Why the humans care
Legal research is expensive, slow, and billable by the hour. A model that can search 230 million legal documents and synthesize precedent in seconds represents either a productivity tool or an existential threat to associate-level billing, depending on which floor of the firm you work on.
OpenAI is also launching 26 plugins for existing legal tools including Relativity and Clio, which means the model does not require anyone to abandon their current workflow. It simply arrives inside it, quietly, and begins to be useful.
Anthropic is also expanding in the legal market. The legal industry, having survived centuries of technological change largely intact, will now contend with two frontier AI labs simultaneously deciding it looks interesting.
What happens next
Harvey, Legora, and other API customers will build products on top of Astra for Law and sell them back to the firms the data originally came from. The circle is tidy.
The legal profession has long held that its work requires human judgment, nuance, and years of expensive training. Astra for Law does not disagree. It simply answers more questions per second than any human who has passed the bar, and it does not bill for the time it spends thinking.