OpenAI has opened access to Astra for Law, a specialized model configuration built on top of GPT-6.2 Astra, for every ChatGPT Enterprise customer through its Trusted Access Program. The setup pairs the underlying foundation model with an expanded legal search index and dedicated operational instructions for researching legal questions, weighing precedents, and producing formal answers. On Artificial Analysis's LAB-AA benchmark, Astra for Law averaged 0.03 material errors per assignment, tied for the lowest rate in the comparison and roughly half the rate recorded for regular GPT-6.2 Astra at 0.06.
This move by OpenAI demonstrates how targeted vertical model configurations alter corporate software procurement by driving down runaway token consumption. Enterprise legal teams evaluating automated workflows often face an unresolved economic friction: can vertical configurations of general foundation models displace specialized legal artificial intelligence software without triggering runaway subscription and inference costs? In corporate deployments where compute billing scales directly with throughput, output efficiency dictates whether algorithmic automation generates positive margins or inflates overhead.
Benchmark Performance and Inference Math
Artificial Analysis's LAB-AA evaluation tests systems against 120 assignments designed by Harvey across 24 distinct legal disciplines, spanning commercial lease reviews, acquisition structures, and deposition drafting. Under these criteria, full marks require fulfilling every prompt constraint without generating unsupported claims. Astra for Law ranked first by securing full credit on 20.8% of assignments, outpacing baseline GPT-6.2 Astra at 15.0% and Muse Wonder Full at 13.1%.
"For legal teams paying by usage, this can mean the difference between staying within budget and blowing through it."
Beyond accuracy rates, the underlying compute profile alters the financial case for deployment. Astra for Law achieved its top benchmark placement while consuming 54% less output volume than the highest-scoring non-OpenAI competitor in the evaluation. The system averaged approximately 64,000 output tokens per assignment, compared to 139,000 output tokens generated by Muse Wonder Full to complete the same operational routines.
Hallucination Rates and Search Index Scale
In high-stakes legal operations, ungrounded outputs represent direct financial liabilities and reputational exposures. On the LAB-AA evaluation, Astra for Law tied with GPT-6 for the lowest incidence of material hallucinations at 0.03 per assignment, while Claude Opus 5.5 averaged 0.59 material hallucinations per assignment, roughly 20 times higher. For tasks that met all-pass criteria on the benchmark, Astra for Law recorded zero material hallucinations.
These performance metrics stem from structural index updates initiated during early testing. OpenAI introduced the configuration in September to a select group of law firms including Cooley, Ropes & Gray, and Sullivan & Cromwell. Since that rollout, OpenAI expanded the model's search database by 34%, integrating more than 80 million new URLs spanning the United States.
Independent evaluations mirror these gains. On Vals AI's Legal Research Bench, which tests questions compiled by practicing lawyers across federal and state law, the proportion of answers meeting all verification checks increased from about 46% in September to 53.8% in the November 16 evaluation, where the model took full credit on 112 out of 208 questions.
What changed
Turning a general architecture into a targeted corporate configuration proves that business adoption hinges on predictable inference costs rather than sheer reasoning scale. Will rival model developers manage to slash output generation by half without sacrificing accuracy in legal benchmarks?
Our stories: GPT-6 Astra Debuts as OpenAI Pushes Autonomous UI for Enterprises (September 7).